g_journal.c revision 163853
1/*- 2 * Copyright (c) 2005-2006 Pawel Jakub Dawidek <pjd@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27#include <sys/cdefs.h> 28__FBSDID("$FreeBSD: head/sys/geom/journal/g_journal.c 163853 2006-11-01 01:19:13Z pjd $"); 29 30#include <sys/param.h> 31#include <sys/systm.h> 32#include <sys/kernel.h> 33#include <sys/module.h> 34#include <sys/limits.h> 35#include <sys/lock.h> 36#include <sys/mutex.h> 37#include <sys/bio.h> 38#include <sys/sysctl.h> 39#include <sys/malloc.h> 40#include <sys/mount.h> 41#include <sys/eventhandler.h> 42#include <sys/proc.h> 43#include <sys/kthread.h> 44#include <sys/sched.h> 45#include <sys/taskqueue.h> 46#include <sys/vnode.h> 47#include <sys/sbuf.h> 48#ifdef GJ_MEMDEBUG 49#include <sys/stack.h> 50#include <sys/kdb.h> 51#endif 52#include <vm/vm.h> 53#include <vm/vm_kern.h> 54#include <geom/geom.h> 55 56#include <geom/journal/g_journal.h> 57 58 59/* 60 * On-disk journal format: 61 * 62 * JH - Journal header 63 * RH - Record header 64 * 65 * %%%%%% ****** +------+ +------+ ****** +------+ %%%%%% 66 * % JH % * RH * | Data | | Data | ... * RH * | Data | ... % JH % ... 67 * %%%%%% ****** +------+ +------+ ****** +------+ %%%%%% 68 * 69 */ 70 71CTASSERT(sizeof(struct g_journal_header) <= 512); 72CTASSERT(sizeof(struct g_journal_record_header) <= 512); 73 74static MALLOC_DEFINE(M_JOURNAL, "journal_data", "GEOM_JOURNAL Data"); 75static struct mtx g_journal_cache_mtx; 76MTX_SYSINIT(g_journal_cache, &g_journal_cache_mtx, "cache usage", MTX_DEF); 77 78const struct g_journal_desc *g_journal_filesystems[] = { 79 &g_journal_ufs, 80 NULL 81}; 82 83SYSCTL_DECL(_kern_geom); 84 85int g_journal_debug = 0; 86TUNABLE_INT("kern.geom.journal.debug", &g_journal_debug); 87static u_int g_journal_switch_time = 10; 88static u_int g_journal_force_switch = 70; 89static u_int g_journal_parallel_flushes = 16; 90static u_int g_journal_parallel_copies = 16; 91static u_int g_journal_accept_immediately = 64; 92static u_int g_journal_record_entries = GJ_RECORD_HEADER_NENTRIES; 93static u_int g_journal_do_optimize = 1; 94 95SYSCTL_NODE(_kern_geom, OID_AUTO, journal, CTLFLAG_RW, 0, "GEOM_JOURNAL stuff"); 96SYSCTL_INT(_kern_geom_journal, OID_AUTO, debug, CTLFLAG_RW, &g_journal_debug, 0, 97 "Debug level"); 98SYSCTL_UINT(_kern_geom_journal, OID_AUTO, switch_time, CTLFLAG_RW, 99 &g_journal_switch_time, 0, "Switch journals every N seconds"); 100SYSCTL_UINT(_kern_geom_journal, OID_AUTO, force_switch, CTLFLAG_RW, 101 &g_journal_force_switch, 0, "Force switch when journal is N%% full"); 102SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_flushes, CTLFLAG_RW, 103 &g_journal_parallel_flushes, 0, 104 "Number of flush I/O requests send in parallel"); 105SYSCTL_UINT(_kern_geom_journal, OID_AUTO, accept_immediately, CTLFLAG_RW, 106 &g_journal_accept_immediately, 0, 107 "Number of I/O requests accepted immediatelly"); 108SYSCTL_UINT(_kern_geom_journal, OID_AUTO, parallel_copies, CTLFLAG_RW, 109 &g_journal_parallel_copies, 0, 110 "Number of copy I/O requests send in parallel"); 111static int 112g_journal_record_entries_sysctl(SYSCTL_HANDLER_ARGS) 113{ 114 u_int entries; 115 int error; 116 117 entries = g_journal_record_entries; 118 error = sysctl_handle_int(oidp, &entries, sizeof(entries), req); 119 if (error != 0 || req->newptr == NULL) 120 return (error); 121 if (entries < 1 || entries > GJ_RECORD_HEADER_NENTRIES) 122 return (EINVAL); 123 g_journal_record_entries = entries; 124 return (0); 125} 126SYSCTL_PROC(_kern_geom_journal, OID_AUTO, record_entries, 127 CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_record_entries_sysctl, "I", 128 "Maximum number of entires in one journal record"); 129SYSCTL_UINT(_kern_geom_journal, OID_AUTO, optimize, CTLFLAG_RW, 130 &g_journal_do_optimize, 0, "Try to combine bios on flush and copy"); 131 132static u_int g_journal_cache_used = 0; 133static u_int g_journal_cache_limit = 64 * 1024 * 1024; 134TUNABLE_INT("kern.geom.journal.cache.limit", &g_journal_cache_limit); 135static u_int g_journal_cache_divisor = 2; 136TUNABLE_INT("kern.geom.journal.cache.divisor", &g_journal_cache_divisor); 137static u_int g_journal_cache_switch = 90; 138static u_int g_journal_cache_misses = 0; 139static u_int g_journal_cache_alloc_failures = 0; 140static u_int g_journal_cache_low = 0; 141 142SYSCTL_NODE(_kern_geom_journal, OID_AUTO, cache, CTLFLAG_RW, 0, 143 "GEOM_JOURNAL cache"); 144SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, used, CTLFLAG_RD, 145 &g_journal_cache_used, 0, "Number of allocated bytes"); 146static int 147g_journal_cache_limit_sysctl(SYSCTL_HANDLER_ARGS) 148{ 149 u_int limit; 150 int error; 151 152 limit = g_journal_cache_limit; 153 error = sysctl_handle_int(oidp, &limit, sizeof(limit), req); 154 if (error != 0 || req->newptr == NULL) 155 return (error); 156 g_journal_cache_limit = limit; 157 g_journal_cache_low = (limit / 100) * g_journal_cache_switch; 158 return (0); 159} 160SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, limit, 161 CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_cache_limit_sysctl, "I", 162 "Maximum number of allocated bytes"); 163SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, divisor, CTLFLAG_RDTUN, 164 &g_journal_cache_divisor, 0, 165 "(kmem_size / kern.geom.journal.cache.divisor) == cache size"); 166static int 167g_journal_cache_switch_sysctl(SYSCTL_HANDLER_ARGS) 168{ 169 u_int cswitch; 170 int error; 171 172 cswitch = g_journal_cache_switch; 173 error = sysctl_handle_int(oidp, &cswitch, sizeof(cswitch), req); 174 if (error != 0 || req->newptr == NULL) 175 return (error); 176 if (cswitch < 0 || cswitch > 100) 177 return (EINVAL); 178 g_journal_cache_switch = cswitch; 179 g_journal_cache_low = (g_journal_cache_limit / 100) * cswitch; 180 return (0); 181} 182SYSCTL_PROC(_kern_geom_journal_cache, OID_AUTO, switch, 183 CTLTYPE_UINT | CTLFLAG_RW, NULL, 0, g_journal_cache_switch_sysctl, "I", 184 "Force switch when we hit this percent of cache use"); 185SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, misses, CTLFLAG_RW, 186 &g_journal_cache_misses, 0, "Number of cache misses"); 187SYSCTL_UINT(_kern_geom_journal_cache, OID_AUTO, alloc_failures, CTLFLAG_RW, 188 &g_journal_cache_alloc_failures, 0, "Memory allocation failures"); 189 190static u_long g_journal_stats_bytes_skipped = 0; 191static u_long g_journal_stats_combined_ios = 0; 192static u_long g_journal_stats_switches = 0; 193static u_long g_journal_stats_wait_for_copy = 0; 194static u_long g_journal_stats_journal_full = 0; 195static u_long g_journal_stats_low_mem = 0; 196 197SYSCTL_NODE(_kern_geom_journal, OID_AUTO, stats, CTLFLAG_RW, 0, 198 "GEOM_JOURNAL statistics"); 199SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, skipped_bytes, CTLFLAG_RW, 200 &g_journal_stats_bytes_skipped, 0, "Number of skipped bytes"); 201SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, combined_ios, CTLFLAG_RW, 202 &g_journal_stats_combined_ios, 0, "Number of combined I/O requests"); 203SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, switches, CTLFLAG_RW, 204 &g_journal_stats_switches, 0, "Number of journal switches"); 205SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, wait_for_copy, CTLFLAG_RW, 206 &g_journal_stats_wait_for_copy, 0, "Wait for journal copy on switch"); 207SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, journal_full, CTLFLAG_RW, 208 &g_journal_stats_journal_full, 0, 209 "Number of times journal was almost full."); 210SYSCTL_ULONG(_kern_geom_journal_stats, OID_AUTO, low_mem, CTLFLAG_RW, 211 &g_journal_stats_low_mem, 0, "Number of times low_mem hook was called."); 212 213static g_taste_t g_journal_taste; 214static g_ctl_req_t g_journal_config; 215static g_dumpconf_t g_journal_dumpconf; 216static g_init_t g_journal_init; 217static g_fini_t g_journal_fini; 218 219struct g_class g_journal_class = { 220 .name = G_JOURNAL_CLASS_NAME, 221 .version = G_VERSION, 222 .taste = g_journal_taste, 223 .ctlreq = g_journal_config, 224 .dumpconf = g_journal_dumpconf, 225 .init = g_journal_init, 226 .fini = g_journal_fini 227}; 228 229static int g_journal_destroy(struct g_journal_softc *sc); 230static void g_journal_metadata_update(struct g_journal_softc *sc); 231static void g_journal_switch_wait(struct g_journal_softc *sc); 232 233#define GJ_SWITCHER_WORKING 0 234#define GJ_SWITCHER_DIE 1 235#define GJ_SWITCHER_DIED 2 236static int g_journal_switcher_state = GJ_SWITCHER_WORKING; 237static int g_journal_switcher_wokenup = 0; 238static int g_journal_sync_requested = 0; 239 240#ifdef GJ_MEMDEBUG 241struct meminfo { 242 size_t mi_size; 243 struct stack mi_stack; 244}; 245#endif 246 247/* 248 * We use our own malloc/realloc/free funtions, so we can collect statistics 249 * and force journal switch when we're running out of cache. 250 */ 251static void * 252gj_malloc(size_t size, int flags) 253{ 254 void *p; 255#ifdef GJ_MEMDEBUG 256 struct meminfo *mi; 257#endif 258 259 mtx_lock(&g_journal_cache_mtx); 260 if (g_journal_cache_limit > 0 && !g_journal_switcher_wokenup && 261 g_journal_cache_used + size > g_journal_cache_low) { 262 GJ_DEBUG(1, "No cache, waking up the switcher."); 263 g_journal_switcher_wokenup = 1; 264 wakeup(&g_journal_switcher_state); 265 } 266 if ((flags & M_NOWAIT) && g_journal_cache_limit > 0 && 267 g_journal_cache_used + size > g_journal_cache_limit) { 268 mtx_unlock(&g_journal_cache_mtx); 269 g_journal_cache_alloc_failures++; 270 return (NULL); 271 } 272 g_journal_cache_used += size; 273 mtx_unlock(&g_journal_cache_mtx); 274 flags &= ~M_NOWAIT; 275#ifndef GJ_MEMDEBUG 276 p = malloc(size, M_JOURNAL, flags | M_WAITOK); 277#else 278 mi = malloc(sizeof(*mi) + size, M_JOURNAL, flags | M_WAITOK); 279 p = (u_char *)mi + sizeof(*mi); 280 mi->mi_size = size; 281 stack_save(&mi->mi_stack); 282#endif 283 return (p); 284} 285 286static void 287gj_free(void *p, size_t size) 288{ 289#ifdef GJ_MEMDEBUG 290 struct meminfo *mi; 291#endif 292 293 KASSERT(p != NULL, ("p=NULL")); 294 KASSERT(size > 0, ("size=0")); 295 mtx_lock(&g_journal_cache_mtx); 296 KASSERT(g_journal_cache_used >= size, ("Freeing too much?")); 297 g_journal_cache_used -= size; 298 mtx_unlock(&g_journal_cache_mtx); 299#ifdef GJ_MEMDEBUG 300 mi = p = (void *)((u_char *)p - sizeof(*mi)); 301 if (mi->mi_size != size) { 302 printf("GJOURNAL: Size mismatch! %zu != %zu\n", size, 303 mi->mi_size); 304 printf("GJOURNAL: Alloc backtrace:\n"); 305 stack_print(&mi->mi_stack); 306 printf("GJOURNAL: Free backtrace:\n"); 307 kdb_backtrace(); 308 } 309#endif 310 free(p, M_JOURNAL); 311} 312 313static void * 314gj_realloc(void *p, size_t size, size_t oldsize) 315{ 316 void *np; 317 318#ifndef GJ_MEMDEBUG 319 mtx_lock(&g_journal_cache_mtx); 320 g_journal_cache_used -= oldsize; 321 g_journal_cache_used += size; 322 mtx_unlock(&g_journal_cache_mtx); 323 np = realloc(p, size, M_JOURNAL, M_WAITOK); 324#else 325 np = gj_malloc(size, M_WAITOK); 326 bcopy(p, np, MIN(oldsize, size)); 327 gj_free(p, oldsize); 328#endif 329 return (np); 330} 331 332static void 333g_journal_check_overflow(struct g_journal_softc *sc) 334{ 335 off_t length, used; 336 337 if ((sc->sc_active.jj_offset < sc->sc_inactive.jj_offset && 338 sc->sc_journal_offset >= sc->sc_inactive.jj_offset) || 339 (sc->sc_active.jj_offset > sc->sc_inactive.jj_offset && 340 sc->sc_journal_offset >= sc->sc_inactive.jj_offset && 341 sc->sc_journal_offset < sc->sc_active.jj_offset)) { 342 panic("Journal overflow (joffset=%jd active=%jd inactive=%jd)", 343 (intmax_t)sc->sc_journal_offset, 344 (intmax_t)sc->sc_active.jj_offset, 345 (intmax_t)sc->sc_inactive.jj_offset); 346 } 347 if (sc->sc_active.jj_offset < sc->sc_inactive.jj_offset) { 348 length = sc->sc_inactive.jj_offset - sc->sc_active.jj_offset; 349 used = sc->sc_journal_offset - sc->sc_active.jj_offset; 350 } else { 351 length = sc->sc_jend - sc->sc_active.jj_offset; 352 length += sc->sc_inactive.jj_offset - sc->sc_jstart; 353 if (sc->sc_journal_offset >= sc->sc_active.jj_offset) 354 used = sc->sc_journal_offset - sc->sc_active.jj_offset; 355 else { 356 used = sc->sc_jend - sc->sc_active.jj_offset; 357 used += sc->sc_journal_offset - sc->sc_jstart; 358 } 359 } 360 /* Already woken up? */ 361 if (g_journal_switcher_wokenup) 362 return; 363 /* 364 * If the active journal takes more than g_journal_force_switch precent 365 * of free journal space, we force journal switch. 366 */ 367 KASSERT(length > 0, 368 ("length=%jd used=%jd active=%jd inactive=%jd joffset=%jd", 369 (intmax_t)length, (intmax_t)used, 370 (intmax_t)sc->sc_active.jj_offset, 371 (intmax_t)sc->sc_inactive.jj_offset, 372 (intmax_t)sc->sc_journal_offset)); 373 if ((used * 100) / length > g_journal_force_switch) { 374 g_journal_stats_journal_full++; 375 GJ_DEBUG(1, "Journal %s %jd%% full, forcing journal switch.", 376 sc->sc_name, (used * 100) / length); 377 mtx_lock(&g_journal_cache_mtx); 378 g_journal_switcher_wokenup = 1; 379 wakeup(&g_journal_switcher_state); 380 mtx_unlock(&g_journal_cache_mtx); 381 } 382} 383 384static void 385g_journal_orphan(struct g_consumer *cp) 386{ 387 struct g_journal_softc *sc; 388 char name[256]; 389 int error; 390 391 g_topology_assert(); 392 sc = cp->geom->softc; 393 GJ_DEBUG(0, "Lost provider %s (journal=%s).", cp->provider->name, 394 sc->sc_name); 395 strlcpy(name, sc->sc_name, sizeof(name)); 396 error = g_journal_destroy(sc); 397 if (error == 0) 398 GJ_DEBUG(0, "Journal %s destroyed.", name); 399 else { 400 GJ_DEBUG(0, "Cannot destroy journal %s (error=%d). " 401 "Destroy it manually after last close.", sc->sc_name, 402 error); 403 } 404} 405 406static int 407g_journal_access(struct g_provider *pp, int acr, int acw, int ace) 408{ 409 struct g_journal_softc *sc; 410 int dcr, dcw, dce; 411 412 g_topology_assert(); 413 GJ_DEBUG(2, "Access request for %s: r%dw%de%d.", pp->name, 414 acr, acw, ace); 415 416 dcr = pp->acr + acr; 417 dcw = pp->acw + acw; 418 dce = pp->ace + ace; 419 420 sc = pp->geom->softc; 421 if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY)) { 422 if (acr <= 0 && acw <= 0 && ace <= 0) 423 return (0); 424 else 425 return (ENXIO); 426 } 427 if (pp->acw == 0 && dcw > 0) { 428 GJ_DEBUG(1, "Marking %s as dirty.", sc->sc_name); 429 sc->sc_flags &= ~GJF_DEVICE_CLEAN; 430 g_topology_unlock(); 431 g_journal_metadata_update(sc); 432 g_topology_lock(); 433 } /* else if (pp->acw == 0 && dcw > 0 && JEMPTY(sc)) { 434 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name); 435 sc->sc_flags |= GJF_DEVICE_CLEAN; 436 g_topology_unlock(); 437 g_journal_metadata_update(sc); 438 g_topology_lock(); 439 } */ 440 return (0); 441} 442 443static void 444g_journal_header_encode(struct g_journal_header *hdr, u_char *data) 445{ 446 447 bcopy(GJ_HEADER_MAGIC, data, sizeof(GJ_HEADER_MAGIC)); 448 data += sizeof(GJ_HEADER_MAGIC); 449 le32enc(data, hdr->jh_journal_id); 450 data += 4; 451 le32enc(data, hdr->jh_journal_next_id); 452} 453 454static int 455g_journal_header_decode(const u_char *data, struct g_journal_header *hdr) 456{ 457 458 bcopy(data, hdr->jh_magic, sizeof(hdr->jh_magic)); 459 data += sizeof(hdr->jh_magic); 460 if (bcmp(hdr->jh_magic, GJ_HEADER_MAGIC, sizeof(GJ_HEADER_MAGIC)) != 0) 461 return (EINVAL); 462 hdr->jh_journal_id = le32dec(data); 463 data += 4; 464 hdr->jh_journal_next_id = le32dec(data); 465 return (0); 466} 467 468static void 469g_journal_flush_cache(struct g_journal_softc *sc) 470{ 471 struct bintime bt; 472 int error; 473 474 if (sc->sc_bio_flush == 0) 475 return; 476 GJ_TIMER_START(1, &bt); 477 if (sc->sc_bio_flush & GJ_FLUSH_JOURNAL) { 478 error = g_io_flush(sc->sc_jconsumer); 479 GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.", 480 sc->sc_jconsumer->provider->name, error); 481 } 482 if (sc->sc_bio_flush & GJ_FLUSH_DATA) { 483 /* 484 * TODO: This could be called in parallel with the 485 * previous call. 486 */ 487 error = g_io_flush(sc->sc_dconsumer); 488 GJ_DEBUG(error == 0 ? 2 : 0, "Flush cache of %s: error=%d.", 489 sc->sc_dconsumer->provider->name, error); 490 } 491 GJ_TIMER_STOP(1, &bt, "Cache flush time"); 492} 493 494static int 495g_journal_write_header(struct g_journal_softc *sc) 496{ 497 struct g_journal_header hdr; 498 struct g_consumer *cp; 499 u_char *buf; 500 int error; 501 502 cp = sc->sc_jconsumer; 503 buf = gj_malloc(cp->provider->sectorsize, M_WAITOK); 504 505 strlcpy(hdr.jh_magic, GJ_HEADER_MAGIC, sizeof(hdr.jh_magic)); 506 hdr.jh_journal_id = sc->sc_journal_id; 507 hdr.jh_journal_next_id = sc->sc_journal_next_id; 508 g_journal_header_encode(&hdr, buf); 509 error = g_write_data(cp, sc->sc_journal_offset, buf, 510 cp->provider->sectorsize); 511 /* if (error == 0) */ 512 sc->sc_journal_offset += cp->provider->sectorsize; 513 514 gj_free(buf, cp->provider->sectorsize); 515 return (error); 516} 517 518/* 519 * Every journal record has a header and data following it. 520 * Functions below are used to decode the header before storing it to 521 * little endian and to encode it after reading to system endianess. 522 */ 523static void 524g_journal_record_header_encode(struct g_journal_record_header *hdr, 525 u_char *data) 526{ 527 struct g_journal_entry *ent; 528 u_int i; 529 530 bcopy(GJ_RECORD_HEADER_MAGIC, data, sizeof(GJ_RECORD_HEADER_MAGIC)); 531 data += sizeof(GJ_RECORD_HEADER_MAGIC); 532 le32enc(data, hdr->jrh_journal_id); 533 data += 8; 534 le16enc(data, hdr->jrh_nentries); 535 data += 2; 536 bcopy(hdr->jrh_sum, data, sizeof(hdr->jrh_sum)); 537 data += 8; 538 for (i = 0; i < hdr->jrh_nentries; i++) { 539 ent = &hdr->jrh_entries[i]; 540 le64enc(data, ent->je_joffset); 541 data += 8; 542 le64enc(data, ent->je_offset); 543 data += 8; 544 le64enc(data, ent->je_length); 545 data += 8; 546 } 547} 548 549static int 550g_journal_record_header_decode(const u_char *data, 551 struct g_journal_record_header *hdr) 552{ 553 struct g_journal_entry *ent; 554 u_int i; 555 556 bcopy(data, hdr->jrh_magic, sizeof(hdr->jrh_magic)); 557 data += sizeof(hdr->jrh_magic); 558 if (strcmp(hdr->jrh_magic, GJ_RECORD_HEADER_MAGIC) != 0) 559 return (EINVAL); 560 hdr->jrh_journal_id = le32dec(data); 561 data += 8; 562 hdr->jrh_nentries = le16dec(data); 563 data += 2; 564 if (hdr->jrh_nentries > GJ_RECORD_HEADER_NENTRIES) 565 return (EINVAL); 566 bcopy(data, hdr->jrh_sum, sizeof(hdr->jrh_sum)); 567 data += 8; 568 for (i = 0; i < hdr->jrh_nentries; i++) { 569 ent = &hdr->jrh_entries[i]; 570 ent->je_joffset = le64dec(data); 571 data += 8; 572 ent->je_offset = le64dec(data); 573 data += 8; 574 ent->je_length = le64dec(data); 575 data += 8; 576 } 577 return (0); 578} 579 580/* 581 * Function reads metadata from a provider (via the given consumer), decodes 582 * it to system endianess and verifies its correctness. 583 */ 584static int 585g_journal_metadata_read(struct g_consumer *cp, struct g_journal_metadata *md) 586{ 587 struct g_provider *pp; 588 u_char *buf; 589 int error; 590 591 g_topology_assert(); 592 593 error = g_access(cp, 1, 0, 0); 594 if (error != 0) 595 return (error); 596 pp = cp->provider; 597 g_topology_unlock(); 598 /* Metadata is stored in last sector. */ 599 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize, 600 &error); 601 g_topology_lock(); 602 g_access(cp, -1, 0, 0); 603 if (error != 0) { 604 GJ_DEBUG(1, "Cannot read metadata from %s (error=%d).", 605 cp->provider->name, error); 606 if (buf != NULL) 607 g_free(buf); 608 return (error); 609 } 610 611 /* Decode metadata. */ 612 error = journal_metadata_decode(buf, md); 613 g_free(buf); 614 /* Is this is gjournal provider at all? */ 615 if (strcmp(md->md_magic, G_JOURNAL_MAGIC) != 0) 616 return (EINVAL); 617 /* 618 * Are we able to handle this version of metadata? 619 * We only maintain backward compatibility. 620 */ 621 if (md->md_version > G_JOURNAL_VERSION) { 622 GJ_DEBUG(0, 623 "Kernel module is too old to handle metadata from %s.", 624 cp->provider->name); 625 return (EINVAL); 626 } 627 /* Is checksum correct? */ 628 if (error != 0) { 629 GJ_DEBUG(0, "MD5 metadata hash mismatch for provider %s.", 630 cp->provider->name); 631 return (error); 632 } 633 return (0); 634} 635 636/* 637 * Two functions below are responsible for updating metadata. 638 * Only metadata on the data provider is updated (we need to update 639 * information about active journal in there). 640 */ 641static void 642g_journal_metadata_done(struct bio *bp) 643{ 644 645 /* 646 * There is not much we can do on error except informing about it. 647 */ 648 if (bp->bio_error != 0) { 649 GJ_LOGREQ(0, bp, "Cannot update metadata (error=%d).", 650 bp->bio_error); 651 } else { 652 GJ_LOGREQ(2, bp, "Metadata updated."); 653 } 654 gj_free(bp->bio_data, bp->bio_length); 655 g_destroy_bio(bp); 656} 657 658static void 659g_journal_metadata_update(struct g_journal_softc *sc) 660{ 661 struct g_journal_metadata md; 662 struct g_consumer *cp; 663 struct bio *bp; 664 u_char *sector; 665 666 cp = sc->sc_dconsumer; 667 sector = gj_malloc(cp->provider->sectorsize, M_WAITOK); 668 strlcpy(md.md_magic, G_JOURNAL_MAGIC, sizeof(md.md_magic)); 669 md.md_version = G_JOURNAL_VERSION; 670 md.md_id = sc->sc_id; 671 md.md_type = sc->sc_orig_type; 672 md.md_jstart = sc->sc_jstart; 673 md.md_jend = sc->sc_jend; 674 md.md_joffset = sc->sc_inactive.jj_offset; 675 md.md_jid = sc->sc_journal_previous_id; 676 md.md_flags = 0; 677 if (sc->sc_flags & GJF_DEVICE_CLEAN) 678 md.md_flags |= GJ_FLAG_CLEAN; 679 680 if (sc->sc_flags & GJF_DEVICE_HARDCODED) 681 strlcpy(md.md_provider, sc->sc_name, sizeof(md.md_provider)); 682 else 683 bzero(md.md_provider, sizeof(md.md_provider)); 684 md.md_provsize = cp->provider->mediasize; 685 journal_metadata_encode(&md, sector); 686 687 /* 688 * Flush the cache, so we know all data are on disk. 689 * We write here informations like "journal is consistent", so we need 690 * to be sure it is. Without BIO_FLUSH here, we can end up in situation 691 * where metadata is stored on disk, but not all data. 692 */ 693 g_journal_flush_cache(sc); 694 695 bp = g_alloc_bio(); 696 bp->bio_offset = cp->provider->mediasize - cp->provider->sectorsize; 697 bp->bio_length = cp->provider->sectorsize; 698 bp->bio_data = sector; 699 bp->bio_cmd = BIO_WRITE; 700 if (!(sc->sc_flags & GJF_DEVICE_DESTROY)) { 701 bp->bio_done = g_journal_metadata_done; 702 g_io_request(bp, cp); 703 } else { 704 bp->bio_done = NULL; 705 g_io_request(bp, cp); 706 biowait(bp, "gjmdu"); 707 g_journal_metadata_done(bp); 708 } 709 710 /* 711 * Be sure metadata reached the disk. 712 */ 713 g_journal_flush_cache(sc); 714} 715 716/* 717 * This is where the I/O request comes from the GEOM. 718 */ 719static void 720g_journal_start(struct bio *bp) 721{ 722 struct g_journal_softc *sc; 723 724 sc = bp->bio_to->geom->softc; 725 GJ_LOGREQ(3, bp, "Request received."); 726 727 switch (bp->bio_cmd) { 728 case BIO_READ: 729 case BIO_WRITE: 730 mtx_lock(&sc->sc_mtx); 731 bioq_insert_tail(&sc->sc_regular_queue, bp); 732 wakeup(sc); 733 mtx_unlock(&sc->sc_mtx); 734 return; 735 case BIO_GETATTR: 736 if (strcmp(bp->bio_attribute, "GJOURNAL::provider") == 0) { 737 strlcpy(bp->bio_data, bp->bio_to->name, bp->bio_length); 738 bp->bio_completed = strlen(bp->bio_to->name) + 1; 739 g_io_deliver(bp, 0); 740 return; 741 } 742 /* FALLTHROUGH */ 743 case BIO_DELETE: 744 default: 745 g_io_deliver(bp, EOPNOTSUPP); 746 return; 747 } 748} 749 750static void 751g_journal_std_done(struct bio *bp) 752{ 753 struct g_journal_softc *sc; 754 755 sc = bp->bio_from->geom->softc; 756 mtx_lock(&sc->sc_mtx); 757 bioq_insert_tail(&sc->sc_back_queue, bp); 758 wakeup(sc); 759 mtx_unlock(&sc->sc_mtx); 760} 761 762static struct bio * 763g_journal_new_bio(off_t start, off_t end, off_t joffset, u_char *data, 764 int flags) 765{ 766 struct bio *bp; 767 768 bp = g_alloc_bio(); 769 bp->bio_offset = start; 770 bp->bio_joffset = joffset; 771 bp->bio_length = end - start; 772 bp->bio_cmd = BIO_WRITE; 773 bp->bio_done = g_journal_std_done; 774 if (data == NULL) 775 bp->bio_data = NULL; 776 else { 777 bp->bio_data = gj_malloc(bp->bio_length, flags); 778 if (bp->bio_data != NULL) 779 bcopy(data, bp->bio_data, bp->bio_length); 780 } 781 return (bp); 782} 783 784#define g_journal_insert_bio(head, bp, flags) \ 785 g_journal_insert((head), (bp)->bio_offset, \ 786 (bp)->bio_offset + (bp)->bio_length, (bp)->bio_joffset, \ 787 (bp)->bio_data, flags) 788/* 789 * The function below does a lot more than just inserting bio to the queue. 790 * It keeps the queue sorted by offset and ensures that there are no doubled 791 * data (it combines bios where ranges overlap). 792 * 793 * The function returns the number of bios inserted (as bio can be splitted). 794 */ 795static int 796g_journal_insert(struct bio **head, off_t nstart, off_t nend, off_t joffset, 797 u_char *data, int flags) 798{ 799 struct bio *nbp, *cbp, *pbp; 800 off_t cstart, cend; 801 u_char *tmpdata; 802 int n; 803 804 GJ_DEBUG(3, "INSERT(%p): (%jd, %jd, %jd)", *head, nstart, nend, 805 joffset); 806 n = 0; 807 pbp = NULL; 808 GJQ_FOREACH(*head, cbp) { 809 cstart = cbp->bio_offset; 810 cend = cbp->bio_offset + cbp->bio_length; 811 812 if (nstart >= cend) { 813 /* 814 * +-------------+ 815 * | | 816 * | current | +-------------+ 817 * | bio | | | 818 * | | | new | 819 * +-------------+ | bio | 820 * | | 821 * +-------------+ 822 */ 823 GJ_DEBUG(3, "INSERT(%p): 1", *head); 824 } else if (nend <= cstart) { 825 /* 826 * +-------------+ 827 * | | 828 * +-------------+ | current | 829 * | | | bio | 830 * | new | | | 831 * | bio | +-------------+ 832 * | | 833 * +-------------+ 834 */ 835 nbp = g_journal_new_bio(nstart, nend, joffset, data, 836 flags); 837 if (pbp == NULL) 838 *head = nbp; 839 else 840 pbp->bio_next = nbp; 841 nbp->bio_next = cbp; 842 n++; 843 GJ_DEBUG(3, "INSERT(%p): 2 (nbp=%p pbp=%p)", *head, nbp, 844 pbp); 845 goto end; 846 } else if (nstart <= cstart && nend >= cend) { 847 /* 848 * +-------------+ +-------------+ 849 * | current bio | | current bio | 850 * +---+-------------+---+ +-------------+---+ 851 * | | | | | | | 852 * | | | | | | | 853 * | +-------------+ | +-------------+ | 854 * | new bio | | new bio | 855 * +---------------------+ +-----------------+ 856 * 857 * +-------------+ +-------------+ 858 * | current bio | | current bio | 859 * +---+-------------+ +-------------+ 860 * | | | | | 861 * | | | | | 862 * | +-------------+ +-------------+ 863 * | new bio | | new bio | 864 * +-----------------+ +-------------+ 865 */ 866 g_journal_stats_bytes_skipped += cbp->bio_length; 867 cbp->bio_offset = nstart; 868 cbp->bio_joffset = joffset; 869 cbp->bio_length = cend - nstart; 870 if (cbp->bio_data != NULL) { 871 gj_free(cbp->bio_data, cend - cstart); 872 cbp->bio_data = NULL; 873 } 874 if (data != NULL) { 875 cbp->bio_data = gj_malloc(cbp->bio_length, 876 flags); 877 if (cbp->bio_data != NULL) { 878 bcopy(data, cbp->bio_data, 879 cbp->bio_length); 880 } 881 data += cend - nstart; 882 } 883 joffset += cend - nstart; 884 nstart = cend; 885 GJ_DEBUG(3, "INSERT(%p): 3 (cbp=%p)", *head, cbp); 886 } else if (nstart > cstart && nend >= cend) { 887 /* 888 * +-----------------+ +-------------+ 889 * | current bio | | current bio | 890 * | +-------------+ | +---------+---+ 891 * | | | | | | | 892 * | | | | | | | 893 * +---+-------------+ +---+---------+ | 894 * | new bio | | new bio | 895 * +-------------+ +-------------+ 896 */ 897 g_journal_stats_bytes_skipped += cend - nstart; 898 nbp = g_journal_new_bio(nstart, cend, joffset, data, 899 flags); 900 nbp->bio_next = cbp->bio_next; 901 cbp->bio_next = nbp; 902 cbp->bio_length = nstart - cstart; 903 if (cbp->bio_data != NULL) { 904 cbp->bio_data = gj_realloc(cbp->bio_data, 905 cbp->bio_length, cend - cstart); 906 } 907 if (data != NULL) 908 data += cend - nstart; 909 joffset += cend - nstart; 910 nstart = cend; 911 n++; 912 GJ_DEBUG(3, "INSERT(%p): 4 (cbp=%p)", *head, cbp); 913 } else if (nstart > cstart && nend < cend) { 914 /* 915 * +---------------------+ 916 * | current bio | 917 * | +-------------+ | 918 * | | | | 919 * | | | | 920 * +---+-------------+---+ 921 * | new bio | 922 * +-------------+ 923 */ 924 g_journal_stats_bytes_skipped += nend - nstart; 925 nbp = g_journal_new_bio(nstart, nend, joffset, data, 926 flags); 927 nbp->bio_next = cbp->bio_next; 928 cbp->bio_next = nbp; 929 if (cbp->bio_data == NULL) 930 tmpdata = NULL; 931 else 932 tmpdata = cbp->bio_data + nend - cstart; 933 nbp = g_journal_new_bio(nend, cend, 934 cbp->bio_joffset + nend - cstart, tmpdata, flags); 935 nbp->bio_next = ((struct bio *)cbp->bio_next)->bio_next; 936 ((struct bio *)cbp->bio_next)->bio_next = nbp; 937 cbp->bio_length = nstart - cstart; 938 if (cbp->bio_data != NULL) { 939 cbp->bio_data = gj_realloc(cbp->bio_data, 940 cbp->bio_length, cend - cstart); 941 } 942 n += 2; 943 GJ_DEBUG(3, "INSERT(%p): 5 (cbp=%p)", *head, cbp); 944 goto end; 945 } else if (nstart <= cstart && nend < cend) { 946 /* 947 * +-----------------+ +-------------+ 948 * | current bio | | current bio | 949 * +-------------+ | +---+---------+ | 950 * | | | | | | | 951 * | | | | | | | 952 * +-------------+---+ | +---------+---+ 953 * | new bio | | new bio | 954 * +-------------+ +-------------+ 955 */ 956 g_journal_stats_bytes_skipped += nend - nstart; 957 nbp = g_journal_new_bio(nstart, nend, joffset, data, 958 flags); 959 if (pbp == NULL) 960 *head = nbp; 961 else 962 pbp->bio_next = nbp; 963 nbp->bio_next = cbp; 964 cbp->bio_offset = nend; 965 cbp->bio_length = cend - nend; 966 cbp->bio_joffset += nend - cstart; 967 tmpdata = cbp->bio_data; 968 if (tmpdata != NULL) { 969 cbp->bio_data = gj_malloc(cbp->bio_length, 970 flags); 971 if (cbp->bio_data != NULL) { 972 bcopy(tmpdata + nend - cstart, 973 cbp->bio_data, cbp->bio_length); 974 } 975 gj_free(tmpdata, cend - cstart); 976 } 977 n++; 978 GJ_DEBUG(3, "INSERT(%p): 6 (cbp=%p)", *head, cbp); 979 goto end; 980 } 981 if (nstart == nend) 982 goto end; 983 pbp = cbp; 984 } 985 nbp = g_journal_new_bio(nstart, nend, joffset, data, flags); 986 if (pbp == NULL) 987 *head = nbp; 988 else 989 pbp->bio_next = nbp; 990 nbp->bio_next = NULL; 991 n++; 992 GJ_DEBUG(3, "INSERT(%p): 8 (nbp=%p pbp=%p)", *head, nbp, pbp); 993end: 994 if (g_journal_debug >= 3) { 995 GJQ_FOREACH(*head, cbp) { 996 GJ_DEBUG(3, "ELEMENT: %p (%jd, %jd, %jd, %p)", cbp, 997 (intmax_t)cbp->bio_offset, 998 (intmax_t)cbp->bio_length, 999 (intmax_t)cbp->bio_joffset, cbp->bio_data); 1000 } 1001 GJ_DEBUG(3, "INSERT(%p): DONE %d", *head, n); 1002 } 1003 sbuf_delete(sb); 1004 return (n); 1005} 1006 1007/* 1008 * The function combines neighbour bios trying to squeeze as much data as 1009 * possible into one bio. 1010 * 1011 * The function returns the number of bios combined (negative value). 1012 */ 1013static int 1014g_journal_optimize(struct bio *head) 1015{ 1016 struct bio *cbp, *pbp; 1017 int n; 1018 1019 n = 0; 1020 pbp = NULL; 1021 GJQ_FOREACH(head, cbp) { 1022 /* Skip bios which has to be read first. */ 1023 if (cbp->bio_data == NULL) { 1024 pbp = NULL; 1025 continue; 1026 } 1027 /* There is no previous bio yet. */ 1028 if (pbp == NULL) { 1029 pbp = cbp; 1030 continue; 1031 } 1032 /* Is this a neighbour bio? */ 1033 if (pbp->bio_offset + pbp->bio_length != cbp->bio_offset) { 1034 /* Be sure that bios queue is sorted. */ 1035 KASSERT(pbp->bio_offset + pbp->bio_length < cbp->bio_offset, 1036 ("poffset=%jd plength=%jd coffset=%jd", 1037 (intmax_t)pbp->bio_offset, 1038 (intmax_t)pbp->bio_length, 1039 (intmax_t)cbp->bio_offset)); 1040 pbp = cbp; 1041 continue; 1042 } 1043 /* Be sure we don't end up with too big bio. */ 1044 if (pbp->bio_length + cbp->bio_length > MAXPHYS) { 1045 pbp = cbp; 1046 continue; 1047 } 1048 /* Ok, we can join bios. */ 1049 GJ_LOGREQ(4, pbp, "Join: "); 1050 GJ_LOGREQ(4, cbp, "and: "); 1051 pbp->bio_data = gj_realloc(pbp->bio_data, 1052 pbp->bio_length + cbp->bio_length, pbp->bio_length); 1053 bcopy(cbp->bio_data, pbp->bio_data + pbp->bio_length, 1054 cbp->bio_length); 1055 gj_free(cbp->bio_data, cbp->bio_length); 1056 pbp->bio_length += cbp->bio_length; 1057 pbp->bio_next = cbp->bio_next; 1058 g_destroy_bio(cbp); 1059 cbp = pbp; 1060 g_journal_stats_combined_ios++; 1061 n--; 1062 GJ_LOGREQ(4, pbp, "Got: "); 1063 } 1064 return (n); 1065} 1066 1067/* 1068 * TODO: Update comment. 1069 * These are functions responsible for copying one portion of data from journal 1070 * to the destination provider. 1071 * The order goes like this: 1072 * 1. Read the header, which contains informations about data blocks 1073 * following it. 1074 * 2. Read the data blocks from the journal. 1075 * 3. Write the data blocks on the data provider. 1076 * 1077 * g_journal_copy_start() 1078 * g_journal_copy_done() - got finished write request, logs potential errors. 1079 */ 1080 1081/* 1082 * When there is no data in cache, this function is used to read it. 1083 */ 1084static void 1085g_journal_read_first(struct g_journal_softc *sc, struct bio *bp) 1086{ 1087 struct bio *cbp; 1088 1089 /* 1090 * We were short in memory, so data was freed. 1091 * In that case we need to read it back from journal. 1092 */ 1093 cbp = g_alloc_bio(); 1094 cbp->bio_cflags = bp->bio_cflags; 1095 cbp->bio_parent = bp; 1096 cbp->bio_offset = bp->bio_joffset; 1097 cbp->bio_length = bp->bio_length; 1098 cbp->bio_data = gj_malloc(bp->bio_length, M_WAITOK); 1099 cbp->bio_cmd = BIO_READ; 1100 cbp->bio_done = g_journal_std_done; 1101 GJ_LOGREQ(4, cbp, "READ FIRST"); 1102 g_io_request(cbp, sc->sc_jconsumer); 1103 g_journal_cache_misses++; 1104} 1105 1106static void 1107g_journal_copy_send(struct g_journal_softc *sc) 1108{ 1109 struct bio *bioq, *bp, *lbp; 1110 1111 bioq = lbp = NULL; 1112 mtx_lock(&sc->sc_mtx); 1113 for (; sc->sc_copy_in_progress < g_journal_parallel_copies;) { 1114 bp = GJQ_FIRST(sc->sc_inactive.jj_queue); 1115 if (bp == NULL) 1116 break; 1117 GJQ_REMOVE(sc->sc_inactive.jj_queue, bp); 1118 sc->sc_copy_in_progress++; 1119 GJQ_INSERT_AFTER(bioq, bp, lbp); 1120 lbp = bp; 1121 } 1122 mtx_unlock(&sc->sc_mtx); 1123 if (g_journal_do_optimize) 1124 sc->sc_copy_in_progress += g_journal_optimize(bioq); 1125 while ((bp = GJQ_FIRST(bioq)) != NULL) { 1126 GJQ_REMOVE(bioq, bp); 1127 GJQ_INSERT_HEAD(sc->sc_copy_queue, bp); 1128 bp->bio_cflags = GJ_BIO_COPY; 1129 if (bp->bio_data == NULL) 1130 g_journal_read_first(sc, bp); 1131 else { 1132 bp->bio_joffset = 0; 1133 GJ_LOGREQ(4, bp, "SEND"); 1134 g_io_request(bp, sc->sc_dconsumer); 1135 } 1136 } 1137} 1138 1139static void 1140g_journal_copy_start(struct g_journal_softc *sc) 1141{ 1142 1143 /* 1144 * Remember in metadata that we're starting to copy journaled data 1145 * to the data provider. 1146 * In case of power failure, we will copy these data once again on boot. 1147 */ 1148 if (!sc->sc_journal_copying) { 1149 sc->sc_journal_copying = 1; 1150 GJ_DEBUG(1, "Starting copy of journal."); 1151 g_journal_metadata_update(sc); 1152 } 1153 g_journal_copy_send(sc); 1154} 1155 1156/* 1157 * Data block has been read from the journal provider. 1158 */ 1159static int 1160g_journal_copy_read_done(struct bio *bp) 1161{ 1162 struct g_journal_softc *sc; 1163 struct g_consumer *cp; 1164 struct bio *pbp; 1165 1166 KASSERT(bp->bio_cflags == GJ_BIO_COPY, 1167 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY)); 1168 1169 sc = bp->bio_from->geom->softc; 1170 pbp = bp->bio_parent; 1171 1172 if (bp->bio_error != 0) { 1173 GJ_DEBUG(0, "Error while reading data from %s (error=%d).", 1174 bp->bio_to->name, bp->bio_error); 1175 /* 1176 * We will not be able to deliver WRITE request as well. 1177 */ 1178 gj_free(bp->bio_data, bp->bio_length); 1179 g_destroy_bio(pbp); 1180 g_destroy_bio(bp); 1181 sc->sc_copy_in_progress--; 1182 return (1); 1183 } 1184 pbp->bio_data = bp->bio_data; 1185 cp = sc->sc_dconsumer; 1186 g_io_request(pbp, cp); 1187 GJ_LOGREQ(4, bp, "READ DONE"); 1188 g_destroy_bio(bp); 1189 return (0); 1190} 1191 1192/* 1193 * Data block has been written to the data provider. 1194 */ 1195static void 1196g_journal_copy_write_done(struct bio *bp) 1197{ 1198 struct g_journal_softc *sc; 1199 1200 KASSERT(bp->bio_cflags == GJ_BIO_COPY, 1201 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_COPY)); 1202 1203 sc = bp->bio_from->geom->softc; 1204 sc->sc_copy_in_progress--; 1205 1206 if (bp->bio_error != 0) { 1207 GJ_LOGREQ(0, bp, "[copy] Error while writting data (error=%d)", 1208 bp->bio_error); 1209 } 1210 GJQ_REMOVE(sc->sc_copy_queue, bp); 1211 gj_free(bp->bio_data, bp->bio_length); 1212 GJ_LOGREQ(4, bp, "DONE"); 1213 g_destroy_bio(bp); 1214 1215 if (sc->sc_copy_in_progress == 0) { 1216 /* 1217 * This was the last write request for this journal. 1218 */ 1219 GJ_DEBUG(1, "Data has been copied."); 1220 sc->sc_journal_copying = 0; 1221 } 1222} 1223 1224static void g_journal_flush_done(struct bio *bp); 1225 1226/* 1227 * Flush one record onto active journal provider. 1228 */ 1229static void 1230g_journal_flush(struct g_journal_softc *sc) 1231{ 1232 struct g_journal_record_header hdr; 1233 struct g_journal_entry *ent; 1234 struct g_provider *pp; 1235 struct bio **bioq; 1236 struct bio *bp, *fbp, *pbp; 1237 off_t joffset, size; 1238 u_char *data, hash[16]; 1239 MD5_CTX ctx; 1240 u_int i; 1241 1242 if (sc->sc_current_count == 0) 1243 return; 1244 1245 size = 0; 1246 pp = sc->sc_jprovider; 1247 GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc); 1248 joffset = sc->sc_journal_offset; 1249 1250 GJ_DEBUG(2, "Storing %d journal entries on %s at %jd.", 1251 sc->sc_current_count, pp->name, (intmax_t)joffset); 1252 1253 /* 1254 * Store 'journal id', so we know to which journal this record belongs. 1255 */ 1256 hdr.jrh_journal_id = sc->sc_journal_id; 1257 /* Could be less than g_journal_record_entries if called due timeout. */ 1258 hdr.jrh_nentries = MIN(sc->sc_current_count, g_journal_record_entries); 1259 strlcpy(hdr.jrh_magic, GJ_RECORD_HEADER_MAGIC, sizeof(hdr.jrh_magic)); 1260 1261 bioq = &sc->sc_active.jj_queue; 1262 pbp = sc->sc_flush_queue; 1263 1264 fbp = g_alloc_bio(); 1265 fbp->bio_parent = NULL; 1266 fbp->bio_cflags = GJ_BIO_JOURNAL; 1267 fbp->bio_offset = -1; 1268 fbp->bio_joffset = joffset; 1269 fbp->bio_length = pp->sectorsize; 1270 fbp->bio_cmd = BIO_WRITE; 1271 fbp->bio_done = g_journal_std_done; 1272 GJQ_INSERT_AFTER(sc->sc_flush_queue, fbp, pbp); 1273 pbp = fbp; 1274 fbp->bio_to = pp; 1275 GJ_LOGREQ(4, fbp, "FLUSH_OUT"); 1276 joffset += pp->sectorsize; 1277 sc->sc_flush_count++; 1278 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) 1279 MD5Init(&ctx); 1280 1281 for (i = 0; i < hdr.jrh_nentries; i++) { 1282 bp = sc->sc_current_queue; 1283 KASSERT(bp != NULL, ("NULL bp")); 1284 bp->bio_to = pp; 1285 GJ_LOGREQ(4, bp, "FLUSHED"); 1286 sc->sc_current_queue = bp->bio_next; 1287 bp->bio_next = NULL; 1288 sc->sc_current_count--; 1289 1290 /* Add to the header. */ 1291 ent = &hdr.jrh_entries[i]; 1292 ent->je_offset = bp->bio_offset; 1293 ent->je_joffset = joffset; 1294 ent->je_length = bp->bio_length; 1295 size += ent->je_length; 1296 1297 data = bp->bio_data; 1298 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) 1299 MD5Update(&ctx, data, ent->je_length); 1300 bzero(bp, sizeof(*bp)); 1301 bp->bio_cflags = GJ_BIO_JOURNAL; 1302 bp->bio_offset = ent->je_offset; 1303 bp->bio_joffset = ent->je_joffset; 1304 bp->bio_length = ent->je_length; 1305 bp->bio_data = data; 1306 bp->bio_cmd = BIO_WRITE; 1307 bp->bio_done = g_journal_std_done; 1308 GJQ_INSERT_AFTER(sc->sc_flush_queue, bp, pbp); 1309 pbp = bp; 1310 bp->bio_to = pp; 1311 GJ_LOGREQ(4, bp, "FLUSH_OUT"); 1312 joffset += bp->bio_length; 1313 sc->sc_flush_count++; 1314 1315 /* 1316 * Add request to the active sc_journal_queue queue. 1317 * This is our cache. After journal switch we don't have to 1318 * read the data from the inactive journal, because we keep 1319 * it in memory. 1320 */ 1321 g_journal_insert(bioq, ent->je_offset, 1322 ent->je_offset + ent->je_length, ent->je_joffset, data, 1323 M_NOWAIT); 1324 } 1325 1326 /* 1327 * After all requests, store valid header. 1328 */ 1329 data = gj_malloc(pp->sectorsize, M_WAITOK); 1330 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) { 1331 MD5Final(hash, &ctx); 1332 bcopy(hash, hdr.jrh_sum, sizeof(hdr.jrh_sum)); 1333 } 1334 g_journal_record_header_encode(&hdr, data); 1335 fbp->bio_data = data; 1336 1337 sc->sc_journal_offset = joffset; 1338 1339 g_journal_check_overflow(sc); 1340} 1341 1342/* 1343 * Flush request finished. 1344 */ 1345static void 1346g_journal_flush_done(struct bio *bp) 1347{ 1348 struct g_journal_softc *sc; 1349 struct g_consumer *cp; 1350 1351 KASSERT((bp->bio_cflags & GJ_BIO_MASK) == GJ_BIO_JOURNAL, 1352 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_JOURNAL)); 1353 1354 cp = bp->bio_from; 1355 sc = cp->geom->softc; 1356 sc->sc_flush_in_progress--; 1357 1358 if (bp->bio_error != 0) { 1359 GJ_LOGREQ(0, bp, "[flush] Error while writting data (error=%d)", 1360 bp->bio_error); 1361 } 1362 gj_free(bp->bio_data, bp->bio_length); 1363 GJ_LOGREQ(4, bp, "DONE"); 1364 g_destroy_bio(bp); 1365} 1366 1367static void g_journal_release_delayed(struct g_journal_softc *sc); 1368 1369static void 1370g_journal_flush_send(struct g_journal_softc *sc) 1371{ 1372 struct g_consumer *cp; 1373 struct bio *bioq, *bp, *lbp; 1374 1375 cp = sc->sc_jconsumer; 1376 bioq = lbp = NULL; 1377 while (sc->sc_flush_in_progress < g_journal_parallel_flushes) { 1378 /* Send one flush requests to the active journal. */ 1379 bp = GJQ_FIRST(sc->sc_flush_queue); 1380 if (bp != NULL) { 1381 GJQ_REMOVE(sc->sc_flush_queue, bp); 1382 sc->sc_flush_count--; 1383 bp->bio_offset = bp->bio_joffset; 1384 bp->bio_joffset = 0; 1385 sc->sc_flush_in_progress++; 1386 GJQ_INSERT_AFTER(bioq, bp, lbp); 1387 lbp = bp; 1388 } 1389 /* Try to release delayed requests. */ 1390 g_journal_release_delayed(sc); 1391 /* If there are no requests to flush, leave. */ 1392 if (GJQ_FIRST(sc->sc_flush_queue) == NULL) 1393 break; 1394 } 1395 if (g_journal_do_optimize) 1396 sc->sc_flush_in_progress += g_journal_optimize(bioq); 1397 while ((bp = GJQ_FIRST(bioq)) != NULL) { 1398 GJQ_REMOVE(bioq, bp); 1399 GJ_LOGREQ(3, bp, "Flush request send"); 1400 g_io_request(bp, cp); 1401 } 1402} 1403 1404static void 1405g_journal_add_current(struct g_journal_softc *sc, struct bio *bp) 1406{ 1407 int n; 1408 1409 GJ_LOGREQ(4, bp, "CURRENT %d", sc->sc_current_count); 1410 n = g_journal_insert_bio(&sc->sc_current_queue, bp, M_WAITOK); 1411 sc->sc_current_count += n; 1412 n = g_journal_optimize(sc->sc_current_queue); 1413 sc->sc_current_count += n; 1414 /* 1415 * For requests which are added to the current queue we deliver 1416 * response immediately. 1417 */ 1418 bp->bio_completed = bp->bio_length; 1419 g_io_deliver(bp, 0); 1420 if (sc->sc_current_count >= g_journal_record_entries) { 1421 /* 1422 * Let's flush one record onto active journal provider. 1423 */ 1424 g_journal_flush(sc); 1425 } 1426} 1427 1428static void 1429g_journal_release_delayed(struct g_journal_softc *sc) 1430{ 1431 struct bio *bp; 1432 1433 for (;;) { 1434 /* The flush queue is full, exit. */ 1435 if (sc->sc_flush_count >= g_journal_accept_immediately) 1436 return; 1437 bp = bioq_takefirst(&sc->sc_delayed_queue); 1438 if (bp == NULL) 1439 return; 1440 sc->sc_delayed_count--; 1441 g_journal_add_current(sc, bp); 1442 } 1443} 1444 1445/* 1446 * Add I/O request to the current queue. If we have enough requests for one 1447 * journal record we flush them onto active journal provider. 1448 */ 1449static void 1450g_journal_add_request(struct g_journal_softc *sc, struct bio *bp) 1451{ 1452 1453 /* 1454 * The flush queue is full, we need to delay the request. 1455 */ 1456 if (sc->sc_delayed_count > 0 || 1457 sc->sc_flush_count >= g_journal_accept_immediately) { 1458 GJ_LOGREQ(4, bp, "DELAYED"); 1459 bioq_insert_tail(&sc->sc_delayed_queue, bp); 1460 sc->sc_delayed_count++; 1461 return; 1462 } 1463 1464 KASSERT(TAILQ_EMPTY(&sc->sc_delayed_queue.queue), 1465 ("DELAYED queue not empty.")); 1466 g_journal_add_current(sc, bp); 1467} 1468 1469static void g_journal_read_done(struct bio *bp); 1470 1471/* 1472 * Try to find requested data in cache. 1473 */ 1474static struct bio * 1475g_journal_read_find(struct bio *head, int sorted, struct bio *pbp, off_t ostart, 1476 off_t oend) 1477{ 1478 off_t cstart, cend; 1479 struct bio *bp; 1480 1481 GJQ_FOREACH(head, bp) { 1482 if (bp->bio_offset == -1) 1483 continue; 1484 cstart = MAX(ostart, bp->bio_offset); 1485 cend = MIN(oend, bp->bio_offset + bp->bio_length); 1486 if (cend <= ostart) 1487 continue; 1488 else if (cstart >= oend) { 1489 if (!sorted) 1490 continue; 1491 else { 1492 bp = NULL; 1493 break; 1494 } 1495 } 1496 if (bp->bio_data == NULL) 1497 break; 1498 GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend, 1499 bp); 1500 bcopy(bp->bio_data + cstart - bp->bio_offset, 1501 pbp->bio_data + cstart - pbp->bio_offset, cend - cstart); 1502 pbp->bio_completed += cend - cstart; 1503 if (pbp->bio_completed == pbp->bio_length) { 1504 /* 1505 * Cool, the whole request was in cache, deliver happy 1506 * message. 1507 */ 1508 g_io_deliver(pbp, 0); 1509 return (pbp); 1510 } 1511 break; 1512 } 1513 return (bp); 1514} 1515 1516/* 1517 * Try to find requested data in cache. 1518 */ 1519static struct bio * 1520g_journal_read_queue_find(struct bio_queue *head, struct bio *pbp, off_t ostart, 1521 off_t oend) 1522{ 1523 off_t cstart, cend; 1524 struct bio *bp; 1525 1526 TAILQ_FOREACH(bp, head, bio_queue) { 1527 cstart = MAX(ostart, bp->bio_offset); 1528 cend = MIN(oend, bp->bio_offset + bp->bio_length); 1529 if (cend <= ostart) 1530 continue; 1531 else if (cstart >= oend) 1532 continue; 1533 KASSERT(bp->bio_data != NULL, 1534 ("%s: bio_data == NULL", __func__)); 1535 GJ_DEBUG(3, "READ(%p): (%jd, %jd) (bp=%p)", head, cstart, cend, 1536 bp); 1537 bcopy(bp->bio_data + cstart - bp->bio_offset, 1538 pbp->bio_data + cstart - pbp->bio_offset, cend - cstart); 1539 pbp->bio_completed += cend - cstart; 1540 if (pbp->bio_completed == pbp->bio_length) { 1541 /* 1542 * Cool, the whole request was in cache, deliver happy 1543 * message. 1544 */ 1545 g_io_deliver(pbp, 0); 1546 return (pbp); 1547 } 1548 break; 1549 } 1550 return (bp); 1551} 1552 1553/* 1554 * This function is used for colecting data on read. 1555 * The complexity is because parts of the data can be stored in four different 1556 * places: 1557 * - in delayed requests 1558 * - in memory - the data not yet send to the active journal provider 1559 * - in requests which are going to be sent to the active journal 1560 * - in the active journal 1561 * - in the inactive journal 1562 * - in the data provider 1563 */ 1564static void 1565g_journal_read(struct g_journal_softc *sc, struct bio *pbp, off_t ostart, 1566 off_t oend) 1567{ 1568 struct bio *bp, *nbp, *head; 1569 off_t cstart, cend; 1570 u_int i, sorted = 0; 1571 1572 GJ_DEBUG(3, "READ: (%jd, %jd)", ostart, oend); 1573 1574 cstart = cend = -1; 1575 bp = NULL; 1576 head = NULL; 1577 for (i = 0; i <= 5; i++) { 1578 switch (i) { 1579 case 0: /* Delayed requests. */ 1580 head = NULL; 1581 sorted = 0; 1582 break; 1583 case 1: /* Not-yet-send data. */ 1584 head = sc->sc_current_queue; 1585 sorted = 1; 1586 break; 1587 case 2: /* In-flight to the active journal. */ 1588 head = sc->sc_flush_queue; 1589 sorted = 0; 1590 break; 1591 case 3: /* Active journal. */ 1592 head = sc->sc_active.jj_queue; 1593 sorted = 1; 1594 break; 1595 case 4: /* Inactive journal. */ 1596 /* 1597 * XXX: Here could be a race with g_journal_lowmem(). 1598 */ 1599 head = sc->sc_inactive.jj_queue; 1600 sorted = 1; 1601 break; 1602 case 5: /* In-flight to the data provider. */ 1603 head = sc->sc_copy_queue; 1604 sorted = 0; 1605 break; 1606 default: 1607 panic("gjournal %s: i=%d", __func__, i); 1608 } 1609 if (i == 0) 1610 bp = g_journal_read_queue_find(&sc->sc_delayed_queue.queue, pbp, ostart, oend); 1611 else 1612 bp = g_journal_read_find(head, sorted, pbp, ostart, oend); 1613 if (bp == pbp) { /* Got the whole request. */ 1614 GJ_DEBUG(2, "Got the whole request from %u.", i); 1615 return; 1616 } else if (bp != NULL) { 1617 cstart = MAX(ostart, bp->bio_offset); 1618 cend = MIN(oend, bp->bio_offset + bp->bio_length); 1619 GJ_DEBUG(2, "Got part of the request from %u (%jd-%jd).", 1620 i, (intmax_t)cstart, (intmax_t)cend); 1621 break; 1622 } 1623 } 1624 if (bp != NULL) { 1625 if (bp->bio_data == NULL) { 1626 nbp = g_clone_bio(pbp); 1627 nbp->bio_cflags = GJ_BIO_READ; 1628 nbp->bio_data = 1629 pbp->bio_data + cstart - pbp->bio_offset; 1630 nbp->bio_offset = 1631 bp->bio_joffset + cstart - bp->bio_offset; 1632 nbp->bio_length = cend - cstart; 1633 nbp->bio_done = g_journal_read_done; 1634 g_io_request(nbp, sc->sc_jconsumer); 1635 } 1636 /* 1637 * If we don't have the whole request yet, call g_journal_read() 1638 * recursively. 1639 */ 1640 if (ostart < cstart) 1641 g_journal_read(sc, pbp, ostart, cstart); 1642 if (oend > cend) 1643 g_journal_read(sc, pbp, cend, oend); 1644 } else { 1645 /* 1646 * No data in memory, no data in journal. 1647 * Its time for asking data provider. 1648 */ 1649 GJ_DEBUG(3, "READ(data): (%jd, %jd)", ostart, oend); 1650 nbp = g_clone_bio(pbp); 1651 nbp->bio_cflags = GJ_BIO_READ; 1652 nbp->bio_data = pbp->bio_data + ostart - pbp->bio_offset; 1653 nbp->bio_offset = ostart; 1654 nbp->bio_length = oend - ostart; 1655 nbp->bio_done = g_journal_read_done; 1656 g_io_request(nbp, sc->sc_dconsumer); 1657 /* We have the whole request, return here. */ 1658 return; 1659 } 1660} 1661 1662/* 1663 * Function responsible for handling finished READ requests. 1664 * Actually, g_std_done() could be used here, the only difference is that we 1665 * log error. 1666 */ 1667static void 1668g_journal_read_done(struct bio *bp) 1669{ 1670 struct bio *pbp; 1671 1672 KASSERT(bp->bio_cflags == GJ_BIO_READ, 1673 ("Invalid bio (%d != %d).", bp->bio_cflags, GJ_BIO_READ)); 1674 1675 pbp = bp->bio_parent; 1676 pbp->bio_inbed++; 1677 pbp->bio_completed += bp->bio_length; 1678 1679 if (bp->bio_error != 0) { 1680 if (pbp->bio_error == 0) 1681 pbp->bio_error = bp->bio_error; 1682 GJ_DEBUG(0, "Error while reading data from %s (error=%d).", 1683 bp->bio_to->name, bp->bio_error); 1684 } 1685 g_destroy_bio(bp); 1686 if (pbp->bio_children == pbp->bio_inbed && 1687 pbp->bio_completed == pbp->bio_length) { 1688 /* We're done. */ 1689 g_io_deliver(pbp, 0); 1690 } 1691} 1692 1693/* 1694 * Deactive current journal and active next one. 1695 */ 1696static void 1697g_journal_switch(struct g_journal_softc *sc) 1698{ 1699 struct g_provider *pp; 1700 1701 if (JEMPTY(sc)) { 1702 GJ_DEBUG(3, "No need for %s switch.", sc->sc_name); 1703 pp = LIST_FIRST(&sc->sc_geom->provider); 1704 if (!(sc->sc_flags & GJF_DEVICE_CLEAN) && pp->acw == 0) { 1705 sc->sc_flags |= GJF_DEVICE_CLEAN; 1706 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name); 1707 g_journal_metadata_update(sc); 1708 } 1709 } else { 1710 GJ_DEBUG(3, "Switching journal %s.", sc->sc_geom->name); 1711 1712 pp = sc->sc_jprovider; 1713 1714 sc->sc_journal_previous_id = sc->sc_journal_id; 1715 1716 sc->sc_journal_id = sc->sc_journal_next_id; 1717 sc->sc_journal_next_id = arc4random(); 1718 1719 GJ_VALIDATE_OFFSET(sc->sc_journal_offset, sc); 1720 1721 g_journal_write_header(sc); 1722 1723 sc->sc_inactive.jj_offset = sc->sc_active.jj_offset; 1724 sc->sc_inactive.jj_queue = sc->sc_active.jj_queue; 1725 1726 sc->sc_active.jj_offset = 1727 sc->sc_journal_offset - pp->sectorsize; 1728 sc->sc_active.jj_queue = NULL; 1729 1730 /* 1731 * Switch is done, start copying data from the (now) inactive 1732 * journal to the data provider. 1733 */ 1734 g_journal_copy_start(sc); 1735 } 1736 mtx_lock(&sc->sc_mtx); 1737 sc->sc_flags &= ~GJF_DEVICE_SWITCH; 1738 mtx_unlock(&sc->sc_mtx); 1739} 1740 1741static void 1742g_journal_initialize(struct g_journal_softc *sc) 1743{ 1744 1745 sc->sc_journal_id = arc4random(); 1746 sc->sc_journal_next_id = arc4random(); 1747 sc->sc_journal_previous_id = sc->sc_journal_id; 1748 sc->sc_journal_offset = sc->sc_jstart; 1749 sc->sc_inactive.jj_offset = sc->sc_jstart; 1750 g_journal_write_header(sc); 1751 sc->sc_active.jj_offset = sc->sc_jstart; 1752} 1753 1754static void 1755g_journal_mark_as_dirty(struct g_journal_softc *sc) 1756{ 1757 const struct g_journal_desc *desc; 1758 int i; 1759 1760 GJ_DEBUG(1, "Marking file system %s as dirty.", sc->sc_name); 1761 for (i = 0; (desc = g_journal_filesystems[i]) != NULL; i++) 1762 desc->jd_dirty(sc->sc_dconsumer); 1763} 1764 1765/* 1766 * Function read record header from the given journal. 1767 * It is very simlar to g_read_data(9), but it doesn't allocate memory for bio 1768 * and data on every call. 1769 */ 1770static int 1771g_journal_sync_read(struct g_consumer *cp, struct bio *bp, off_t offset, 1772 void *data) 1773{ 1774 int error; 1775 1776 bzero(bp, sizeof(*bp)); 1777 bp->bio_cmd = BIO_READ; 1778 bp->bio_done = NULL; 1779 bp->bio_offset = offset; 1780 bp->bio_length = cp->provider->sectorsize; 1781 bp->bio_data = data; 1782 g_io_request(bp, cp); 1783 error = biowait(bp, "gjs_read"); 1784 return (error); 1785} 1786 1787#if 0 1788/* 1789 * Function is called when we start the journal device and we detect that 1790 * one of the journals was not fully copied. 1791 * The purpose of this function is to read all records headers from journal 1792 * and placed them in the inactive queue, so we can start journal 1793 * synchronization process and the journal provider itself. 1794 * Design decision was taken to not synchronize the whole journal here as it 1795 * can take too much time. Reading headers only and delaying synchronization 1796 * process until after journal provider is started should be the best choice. 1797 */ 1798#endif 1799 1800static void 1801g_journal_sync(struct g_journal_softc *sc) 1802{ 1803 struct g_journal_record_header rhdr; 1804 struct g_journal_entry *ent; 1805 struct g_journal_header jhdr; 1806 struct g_consumer *cp; 1807 struct bio *bp, *fbp, *tbp; 1808 off_t joffset, offset; 1809 u_char *buf, sum[16]; 1810 uint64_t id; 1811 MD5_CTX ctx; 1812 int error, found, i; 1813 1814 found = 0; 1815 fbp = NULL; 1816 cp = sc->sc_jconsumer; 1817 bp = g_alloc_bio(); 1818 buf = gj_malloc(cp->provider->sectorsize, M_WAITOK); 1819 offset = joffset = sc->sc_inactive.jj_offset = sc->sc_journal_offset; 1820 1821 GJ_DEBUG(2, "Looking for termination at %jd.", (intmax_t)joffset); 1822 1823 /* 1824 * Read and decode first journal header. 1825 */ 1826 error = g_journal_sync_read(cp, bp, offset, buf); 1827 if (error != 0) { 1828 GJ_DEBUG(0, "Error while reading journal header from %s.", 1829 cp->provider->name); 1830 goto end; 1831 } 1832 error = g_journal_header_decode(buf, &jhdr); 1833 if (error != 0) { 1834 GJ_DEBUG(0, "Cannot decode journal header from %s.", 1835 cp->provider->name); 1836 goto end; 1837 } 1838 id = sc->sc_journal_id; 1839 if (jhdr.jh_journal_id != sc->sc_journal_id) { 1840 GJ_DEBUG(1, "Journal ID mismatch at %jd (0x%08x != 0x%08x).", 1841 (intmax_t)offset, (u_int)jhdr.jh_journal_id, (u_int)id); 1842 goto end; 1843 } 1844 offset += cp->provider->sectorsize; 1845 id = sc->sc_journal_next_id = jhdr.jh_journal_next_id; 1846 1847 for (;;) { 1848 /* 1849 * If the biggest record won't fit, look for a record header or 1850 * journal header from the begining. 1851 */ 1852 GJ_VALIDATE_OFFSET(offset, sc); 1853 error = g_journal_sync_read(cp, bp, offset, buf); 1854 if (error != 0) { 1855 /* 1856 * Not good. Having an error while reading header 1857 * means, that we cannot read next headers and in 1858 * consequence we cannot find termination. 1859 */ 1860 GJ_DEBUG(0, 1861 "Error while reading record header from %s.", 1862 cp->provider->name); 1863 break; 1864 } 1865 1866 error = g_journal_record_header_decode(buf, &rhdr); 1867 if (error != 0) { 1868 GJ_DEBUG(2, "Not a record header at %jd (error=%d).", 1869 (intmax_t)offset, error); 1870 /* 1871 * This is not a record header. 1872 * If we are lucky, this is next journal header. 1873 */ 1874 error = g_journal_header_decode(buf, &jhdr); 1875 if (error != 0) { 1876 GJ_DEBUG(1, "Not a journal header at %jd (error=%d).", 1877 (intmax_t)offset, error); 1878 /* 1879 * Nope, this is not journal header, which 1880 * bascially means that journal is not 1881 * terminated properly. 1882 */ 1883 error = ENOENT; 1884 break; 1885 } 1886 /* 1887 * Ok. This is header of _some_ journal. Now we need to 1888 * verify if this is header of the _next_ journal. 1889 */ 1890 if (jhdr.jh_journal_id != id) { 1891 GJ_DEBUG(1, "Journal ID mismatch at %jd " 1892 "(0x%08x != 0x%08x).", (intmax_t)offset, 1893 (u_int)jhdr.jh_journal_id, (u_int)id); 1894 error = ENOENT; 1895 break; 1896 } 1897 1898 /* Found termination. */ 1899 found++; 1900 GJ_DEBUG(1, "Found termination at %jd (id=0x%08x).", 1901 (intmax_t)offset, (u_int)id); 1902 sc->sc_active.jj_offset = offset; 1903 sc->sc_journal_offset = 1904 offset + cp->provider->sectorsize; 1905 sc->sc_journal_id = id; 1906 id = sc->sc_journal_next_id = jhdr.jh_journal_next_id; 1907 1908 while ((tbp = fbp) != NULL) { 1909 fbp = tbp->bio_next; 1910 GJ_LOGREQ(3, tbp, "Adding request."); 1911 g_journal_insert_bio(&sc->sc_inactive.jj_queue, 1912 tbp, M_WAITOK); 1913 } 1914 1915 /* Skip journal's header. */ 1916 offset += cp->provider->sectorsize; 1917 continue; 1918 } 1919 1920 /* Skip record's header. */ 1921 offset += cp->provider->sectorsize; 1922 1923 /* 1924 * Add information about every record entry to the inactive 1925 * queue. 1926 */ 1927 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) 1928 MD5Init(&ctx); 1929 for (i = 0; i < rhdr.jrh_nentries; i++) { 1930 ent = &rhdr.jrh_entries[i]; 1931 GJ_DEBUG(3, "Insert entry: %jd %jd.", 1932 (intmax_t)ent->je_offset, (intmax_t)ent->je_length); 1933 g_journal_insert(&fbp, ent->je_offset, 1934 ent->je_offset + ent->je_length, ent->je_joffset, 1935 NULL, M_WAITOK); 1936 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) { 1937 u_char *buf2; 1938 1939 /* 1940 * TODO: Should use faster function (like 1941 * g_journal_sync_read()). 1942 */ 1943 buf2 = g_read_data(cp, offset, ent->je_length, 1944 NULL); 1945 if (buf2 == NULL) 1946 GJ_DEBUG(0, "Cannot read data at %jd.", 1947 (intmax_t)offset); 1948 else { 1949 MD5Update(&ctx, buf2, ent->je_length); 1950 g_free(buf2); 1951 } 1952 } 1953 /* Skip entry's data. */ 1954 offset += ent->je_length; 1955 } 1956 if (sc->sc_flags & GJF_DEVICE_CHECKSUM) { 1957 MD5Final(sum, &ctx); 1958 if (bcmp(sum, rhdr.jrh_sum, sizeof(rhdr.jrh_sum)) != 0) { 1959 GJ_DEBUG(0, "MD5 hash mismatch at %jd!", 1960 (intmax_t)offset); 1961 } 1962 } 1963 } 1964end: 1965 gj_free(bp->bio_data, cp->provider->sectorsize); 1966 g_destroy_bio(bp); 1967 1968 /* Remove bios from unterminated journal. */ 1969 while ((tbp = fbp) != NULL) { 1970 fbp = tbp->bio_next; 1971 g_destroy_bio(tbp); 1972 } 1973 1974 if (found < 1 && joffset > 0) { 1975 GJ_DEBUG(0, "Journal on %s is broken/corrupted. Initializing.", 1976 sc->sc_name); 1977 while ((tbp = sc->sc_inactive.jj_queue) != NULL) { 1978 sc->sc_inactive.jj_queue = tbp->bio_next; 1979 g_destroy_bio(tbp); 1980 } 1981 g_journal_initialize(sc); 1982 g_journal_mark_as_dirty(sc); 1983 } else { 1984 GJ_DEBUG(0, "Journal %s consistent.", sc->sc_name); 1985 g_journal_copy_start(sc); 1986 } 1987} 1988 1989/* 1990 * Wait for requests. 1991 * If we have requests in the current queue, flush them after 3 seconds from the 1992 * last flush. In this way we don't wait forever (or for journal switch) with 1993 * storing not full records on journal. 1994 */ 1995static void 1996g_journal_wait(struct g_journal_softc *sc, time_t last_write) 1997{ 1998 int error, timeout; 1999 2000 GJ_DEBUG(3, "%s: enter", __func__); 2001 if (sc->sc_current_count == 0) { 2002 if (g_journal_debug < 2) 2003 msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", 0); 2004 else { 2005 /* 2006 * If we have debug turned on, show number of elements 2007 * in various queues. 2008 */ 2009 for (;;) { 2010 error = msleep(sc, &sc->sc_mtx, PRIBIO, 2011 "gj:work", hz * 3); 2012 if (error == 0) { 2013 mtx_unlock(&sc->sc_mtx); 2014 break; 2015 } 2016 GJ_DEBUG(3, "Report: current count=%d", 2017 sc->sc_current_count); 2018 GJ_DEBUG(3, "Report: flush count=%d", 2019 sc->sc_flush_count); 2020 GJ_DEBUG(3, "Report: flush in progress=%d", 2021 sc->sc_flush_in_progress); 2022 GJ_DEBUG(3, "Report: copy in progress=%d", 2023 sc->sc_copy_in_progress); 2024 GJ_DEBUG(3, "Report: delayed=%d", 2025 sc->sc_delayed_count); 2026 } 2027 } 2028 GJ_DEBUG(3, "%s: exit 1", __func__); 2029 return; 2030 } 2031 2032 /* 2033 * Flush even not full records every 3 seconds. 2034 */ 2035 timeout = (last_write + 3 - time_second) * hz; 2036 if (timeout <= 0) { 2037 mtx_unlock(&sc->sc_mtx); 2038 g_journal_flush(sc); 2039 g_journal_flush_send(sc); 2040 GJ_DEBUG(3, "%s: exit 2", __func__); 2041 return; 2042 } 2043 error = msleep(sc, &sc->sc_mtx, PRIBIO | PDROP, "gj:work", timeout); 2044 if (error == EWOULDBLOCK) 2045 g_journal_flush_send(sc); 2046 GJ_DEBUG(3, "%s: exit 3", __func__); 2047} 2048 2049/* 2050 * Worker thread. 2051 */ 2052static void 2053g_journal_worker(void *arg) 2054{ 2055 struct g_journal_softc *sc; 2056 struct g_geom *gp; 2057 struct g_provider *pp; 2058 struct bio *bp; 2059 time_t last_write; 2060 int type; 2061 2062 mtx_lock_spin(&sched_lock); 2063 sched_prio(curthread, PRIBIO); 2064 mtx_unlock_spin(&sched_lock); 2065 2066 sc = arg; 2067 2068 if (sc->sc_flags & GJF_DEVICE_CLEAN) { 2069 GJ_DEBUG(0, "Journal %s clean.", sc->sc_name); 2070 g_journal_initialize(sc); 2071 } else { 2072 g_journal_sync(sc); 2073 } 2074 /* 2075 * Check if we can use BIO_FLUSH. 2076 */ 2077 sc->sc_bio_flush = 0; 2078 if (g_io_flush(sc->sc_jconsumer) == 0) { 2079 sc->sc_bio_flush |= GJ_FLUSH_JOURNAL; 2080 GJ_DEBUG(1, "BIO_FLUSH supported by %s.", 2081 sc->sc_jconsumer->provider->name); 2082 } else { 2083 GJ_DEBUG(0, "BIO_FLUSH not supported by %s.", 2084 sc->sc_jconsumer->provider->name); 2085 } 2086 if (sc->sc_jconsumer != sc->sc_dconsumer) { 2087 if (g_io_flush(sc->sc_dconsumer) == 0) { 2088 sc->sc_bio_flush |= GJ_FLUSH_DATA; 2089 GJ_DEBUG(1, "BIO_FLUSH supported by %s.", 2090 sc->sc_dconsumer->provider->name); 2091 } else { 2092 GJ_DEBUG(0, "BIO_FLUSH not supported by %s.", 2093 sc->sc_dconsumer->provider->name); 2094 } 2095 } 2096 2097 gp = sc->sc_geom; 2098 g_topology_lock(); 2099 pp = g_new_providerf(gp, "%s.journal", sc->sc_name); 2100 KASSERT(pp != NULL, ("Cannot create %s.journal.", sc->sc_name)); 2101 pp->mediasize = sc->sc_mediasize; 2102 /* 2103 * There could be a problem when data provider and journal providers 2104 * have different sectorsize, but such scenario is prevented on journal 2105 * creation. 2106 */ 2107 pp->sectorsize = sc->sc_sectorsize; 2108 g_error_provider(pp, 0); 2109 g_topology_unlock(); 2110 last_write = time_second; 2111 2112 for (;;) { 2113 /* Get first request from the queue. */ 2114 mtx_lock(&sc->sc_mtx); 2115 bp = bioq_first(&sc->sc_back_queue); 2116 if (bp != NULL) 2117 type = (bp->bio_cflags & GJ_BIO_MASK); 2118 if (bp == NULL) { 2119 bp = bioq_first(&sc->sc_regular_queue); 2120 if (bp != NULL) 2121 type = GJ_BIO_REGULAR; 2122 } 2123 if (bp == NULL) { 2124try_switch: 2125 if ((sc->sc_flags & GJF_DEVICE_SWITCH) || 2126 (sc->sc_flags & GJF_DEVICE_DESTROY)) { 2127 if (sc->sc_current_count > 0) { 2128 mtx_unlock(&sc->sc_mtx); 2129 g_journal_flush(sc); 2130 g_journal_flush_send(sc); 2131 continue; 2132 } 2133 if (sc->sc_flush_in_progress > 0) 2134 goto sleep; 2135 if (sc->sc_copy_in_progress > 0) 2136 goto sleep; 2137 } 2138 if (sc->sc_flags & GJF_DEVICE_SWITCH) { 2139 mtx_unlock(&sc->sc_mtx); 2140 g_journal_switch(sc); 2141 wakeup(&sc->sc_journal_copying); 2142 continue; 2143 } 2144 if (sc->sc_flags & GJF_DEVICE_DESTROY) { 2145 GJ_DEBUG(1, "Shutting down worker " 2146 "thread for %s.", gp->name); 2147 sc->sc_worker = NULL; 2148 wakeup(&sc->sc_worker); 2149 mtx_unlock(&sc->sc_mtx); 2150 kthread_exit(0); 2151 } 2152sleep: 2153 g_journal_wait(sc, last_write); 2154 continue; 2155 } 2156 /* 2157 * If we're in switch process, we need to delay all new 2158 * write requests until its done. 2159 */ 2160 if ((sc->sc_flags & GJF_DEVICE_SWITCH) && 2161 type == GJ_BIO_REGULAR && bp->bio_cmd == BIO_WRITE) { 2162 GJ_LOGREQ(2, bp, "WRITE on SWITCH"); 2163 goto try_switch; 2164 } 2165 if (type == GJ_BIO_REGULAR) 2166 bioq_remove(&sc->sc_regular_queue, bp); 2167 else 2168 bioq_remove(&sc->sc_back_queue, bp); 2169 mtx_unlock(&sc->sc_mtx); 2170 switch (type) { 2171 case GJ_BIO_REGULAR: 2172 /* Regular request. */ 2173 switch (bp->bio_cmd) { 2174 case BIO_READ: 2175 g_journal_read(sc, bp, bp->bio_offset, 2176 bp->bio_offset + bp->bio_length); 2177 break; 2178 case BIO_WRITE: 2179 last_write = time_second; 2180 g_journal_add_request(sc, bp); 2181 g_journal_flush_send(sc); 2182 break; 2183 default: 2184 panic("Invalid bio_cmd (%d).", bp->bio_cmd); 2185 } 2186 break; 2187 case GJ_BIO_COPY: 2188 switch (bp->bio_cmd) { 2189 case BIO_READ: 2190 if (g_journal_copy_read_done(bp)) 2191 g_journal_copy_send(sc); 2192 break; 2193 case BIO_WRITE: 2194 g_journal_copy_write_done(bp); 2195 g_journal_copy_send(sc); 2196 break; 2197 default: 2198 panic("Invalid bio_cmd (%d).", bp->bio_cmd); 2199 } 2200 break; 2201 case GJ_BIO_JOURNAL: 2202 g_journal_flush_done(bp); 2203 g_journal_flush_send(sc); 2204 break; 2205 case GJ_BIO_READ: 2206 default: 2207 panic("Invalid bio (%d).", type); 2208 } 2209 } 2210} 2211 2212static void 2213g_journal_destroy_event(void *arg, int flags __unused) 2214{ 2215 struct g_journal_softc *sc; 2216 2217 g_topology_assert(); 2218 sc = arg; 2219 g_journal_destroy(sc); 2220} 2221 2222static void 2223g_journal_timeout(void *arg) 2224{ 2225 struct g_journal_softc *sc; 2226 2227 sc = arg; 2228 GJ_DEBUG(0, "Timeout. Journal %s cannot be completed.", 2229 sc->sc_geom->name); 2230 g_post_event(g_journal_destroy_event, sc, M_NOWAIT, NULL); 2231} 2232 2233static struct g_geom * 2234g_journal_create(struct g_class *mp, struct g_provider *pp, 2235 const struct g_journal_metadata *md) 2236{ 2237 struct g_journal_softc *sc; 2238 struct g_geom *gp; 2239 struct g_consumer *cp; 2240 int error; 2241 2242 g_topology_assert(); 2243 /* 2244 * There are two possibilities: 2245 * 1. Data and both journals are on the same provider. 2246 * 2. Data and journals are all on separated providers. 2247 */ 2248 /* Look for journal device with the same ID. */ 2249 LIST_FOREACH(gp, &mp->geom, geom) { 2250 sc = gp->softc; 2251 if (sc == NULL) 2252 continue; 2253 if (sc->sc_id == md->md_id) 2254 break; 2255 } 2256 if (gp == NULL) 2257 sc = NULL; 2258 else if (sc != NULL && (sc->sc_type & md->md_type) != 0) { 2259 GJ_DEBUG(1, "Journal device %u already configured.", sc->sc_id); 2260 return (NULL); 2261 } 2262 if (md->md_type == 0 || (md->md_type & ~GJ_TYPE_COMPLETE) != 0) { 2263 GJ_DEBUG(0, "Invalid type on %s.", pp->name); 2264 return (NULL); 2265 } 2266 if (md->md_type & GJ_TYPE_DATA) { 2267 GJ_DEBUG(0, "Journal %u: %s contains data.", md->md_id, 2268 pp->name); 2269 } 2270 if (md->md_type & GJ_TYPE_JOURNAL) { 2271 GJ_DEBUG(0, "Journal %u: %s contains journal.", md->md_id, 2272 pp->name); 2273 } 2274 2275 if (sc == NULL) { 2276 /* Action geom. */ 2277 sc = malloc(sizeof(*sc), M_JOURNAL, M_WAITOK | M_ZERO); 2278 sc->sc_id = md->md_id; 2279 sc->sc_type = 0; 2280 sc->sc_flags = 0; 2281 sc->sc_worker = NULL; 2282 2283 gp = g_new_geomf(mp, "gjournal %u", sc->sc_id); 2284 gp->start = g_journal_start; 2285 gp->orphan = g_journal_orphan; 2286 gp->access = g_journal_access; 2287 gp->softc = sc; 2288 sc->sc_geom = gp; 2289 2290 mtx_init(&sc->sc_mtx, "gjournal", NULL, MTX_DEF); 2291 2292 bioq_init(&sc->sc_back_queue); 2293 bioq_init(&sc->sc_regular_queue); 2294 bioq_init(&sc->sc_delayed_queue); 2295 sc->sc_delayed_count = 0; 2296 sc->sc_current_queue = NULL; 2297 sc->sc_current_count = 0; 2298 sc->sc_flush_queue = NULL; 2299 sc->sc_flush_count = 0; 2300 sc->sc_flush_in_progress = 0; 2301 sc->sc_copy_queue = NULL; 2302 sc->sc_copy_in_progress = 0; 2303 sc->sc_inactive.jj_queue = NULL; 2304 sc->sc_active.jj_queue = NULL; 2305 2306 callout_init(&sc->sc_callout, CALLOUT_MPSAFE); 2307 if (md->md_type != GJ_TYPE_COMPLETE) { 2308 /* 2309 * Journal and data are on separate providers. 2310 * At this point we have only one of them. 2311 * We setup a timeout in case the other part will not 2312 * appear, so we won't wait forever. 2313 */ 2314 callout_reset(&sc->sc_callout, 5 * hz, 2315 g_journal_timeout, sc); 2316 } 2317 } 2318 2319 /* Remember type of the data provider. */ 2320 if (md->md_type & GJ_TYPE_DATA) 2321 sc->sc_orig_type = md->md_type; 2322 sc->sc_type |= md->md_type; 2323 cp = NULL; 2324 2325 if (md->md_type & GJ_TYPE_DATA) { 2326 if (md->md_flags & GJ_FLAG_CLEAN) 2327 sc->sc_flags |= GJF_DEVICE_CLEAN; 2328 if (md->md_flags & GJ_FLAG_CHECKSUM) 2329 sc->sc_flags |= GJF_DEVICE_CHECKSUM; 2330 cp = g_new_consumer(gp); 2331 error = g_attach(cp, pp); 2332 KASSERT(error == 0, ("Cannot attach to %s (error=%d).", 2333 pp->name, error)); 2334 error = g_access(cp, 1, 1, 1); 2335 if (error != 0) { 2336 GJ_DEBUG(0, "Cannot access %s (error=%d).", pp->name, 2337 error); 2338 g_journal_destroy(sc); 2339 return (NULL); 2340 } 2341 sc->sc_dconsumer = cp; 2342 sc->sc_mediasize = pp->mediasize - pp->sectorsize; 2343 sc->sc_sectorsize = pp->sectorsize; 2344 sc->sc_jstart = md->md_jstart; 2345 sc->sc_jend = md->md_jend; 2346 if (md->md_provider[0] != '\0') 2347 sc->sc_flags |= GJF_DEVICE_HARDCODED; 2348 sc->sc_journal_offset = md->md_joffset; 2349 sc->sc_journal_id = md->md_jid; 2350 sc->sc_journal_previous_id = md->md_jid; 2351 } 2352 if (md->md_type & GJ_TYPE_JOURNAL) { 2353 if (cp == NULL) { 2354 cp = g_new_consumer(gp); 2355 error = g_attach(cp, pp); 2356 KASSERT(error == 0, ("Cannot attach to %s (error=%d).", 2357 pp->name, error)); 2358 error = g_access(cp, 1, 1, 1); 2359 if (error != 0) { 2360 GJ_DEBUG(0, "Cannot access %s (error=%d).", 2361 pp->name, error); 2362 g_journal_destroy(sc); 2363 return (NULL); 2364 } 2365 } else { 2366 /* 2367 * Journal is on the same provider as data, which means 2368 * that data provider ends where journal starts. 2369 */ 2370 sc->sc_mediasize = md->md_jstart; 2371 } 2372 sc->sc_jconsumer = cp; 2373 } 2374 2375 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) { 2376 /* Journal is not complete yet. */ 2377 return (gp); 2378 } else { 2379 /* Journal complete, cancel timeout. */ 2380 callout_drain(&sc->sc_callout); 2381 } 2382 2383 error = kthread_create(g_journal_worker, sc, &sc->sc_worker, 0, 0, 2384 "g_journal %s", sc->sc_name); 2385 if (error != 0) { 2386 GJ_DEBUG(0, "Cannot create worker thread for %s.journal.", 2387 sc->sc_name); 2388 g_journal_destroy(sc); 2389 return (NULL); 2390 } 2391 2392 return (gp); 2393} 2394 2395static void 2396g_journal_destroy_consumer(void *arg, int flags __unused) 2397{ 2398 struct g_consumer *cp; 2399 2400 g_topology_assert(); 2401 cp = arg; 2402 g_detach(cp); 2403 g_destroy_consumer(cp); 2404} 2405 2406static int 2407g_journal_destroy(struct g_journal_softc *sc) 2408{ 2409 struct g_geom *gp; 2410 struct g_provider *pp; 2411 struct g_consumer *cp; 2412 2413 g_topology_assert(); 2414 2415 if (sc == NULL) 2416 return (ENXIO); 2417 2418 gp = sc->sc_geom; 2419 pp = LIST_FIRST(&gp->provider); 2420 if (pp != NULL) { 2421 if (pp->acr != 0 || pp->acw != 0 || pp->ace != 0) { 2422 GJ_DEBUG(1, "Device %s is still open (r%dw%de%d).", 2423 pp->name, pp->acr, pp->acw, pp->ace); 2424 return (EBUSY); 2425 } 2426 g_error_provider(pp, ENXIO); 2427 2428 g_journal_flush(sc); 2429 g_journal_flush_send(sc); 2430 g_journal_switch(sc); 2431 } 2432 2433 sc->sc_flags |= (GJF_DEVICE_DESTROY | GJF_DEVICE_CLEAN); 2434 2435 g_topology_unlock(); 2436 callout_drain(&sc->sc_callout); 2437 mtx_lock(&sc->sc_mtx); 2438 wakeup(sc); 2439 while (sc->sc_worker != NULL) 2440 msleep(&sc->sc_worker, &sc->sc_mtx, PRIBIO, "gj:destroy", 0); 2441 mtx_unlock(&sc->sc_mtx); 2442 2443 if (pp != NULL) { 2444 GJ_DEBUG(1, "Marking %s as clean.", sc->sc_name); 2445 g_journal_metadata_update(sc); 2446 g_topology_lock(); 2447 pp->flags |= G_PF_WITHER; 2448 g_orphan_provider(pp, ENXIO); 2449 } else { 2450 g_topology_lock(); 2451 } 2452 mtx_destroy(&sc->sc_mtx); 2453 2454 if (sc->sc_current_count != 0) { 2455 GJ_DEBUG(0, "Warning! Number of current requests %d.", 2456 sc->sc_current_count); 2457 } 2458 2459 LIST_FOREACH(cp, &gp->consumer, consumer) { 2460 if (cp->acr + cp->acw + cp->ace > 0) 2461 g_access(cp, -1, -1, -1); 2462 /* 2463 * We keep all consumers open for writting, so if I'll detach 2464 * and destroy consumer here, I'll get providers for taste, so 2465 * journal will be started again. 2466 * Sending an event here, prevents this from happening. 2467 */ 2468 g_post_event(g_journal_destroy_consumer, cp, M_WAITOK, NULL); 2469 } 2470 gp->softc = NULL; 2471 g_wither_geom(gp, ENXIO); 2472 free(sc, M_JOURNAL); 2473 return (0); 2474} 2475 2476static void 2477g_journal_taste_orphan(struct g_consumer *cp) 2478{ 2479 2480 KASSERT(1 == 0, ("%s called while tasting %s.", __func__, 2481 cp->provider->name)); 2482} 2483 2484static struct g_geom * 2485g_journal_taste(struct g_class *mp, struct g_provider *pp, int flags __unused) 2486{ 2487 struct g_journal_metadata md; 2488 struct g_consumer *cp; 2489 struct g_geom *gp; 2490 int error; 2491 2492 g_topology_assert(); 2493 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name); 2494 GJ_DEBUG(2, "Tasting %s.", pp->name); 2495 if (pp->geom->class == mp) 2496 return (NULL); 2497 2498 gp = g_new_geomf(mp, "journal:taste"); 2499 /* This orphan function should be never called. */ 2500 gp->orphan = g_journal_taste_orphan; 2501 cp = g_new_consumer(gp); 2502 g_attach(cp, pp); 2503 error = g_journal_metadata_read(cp, &md); 2504 g_detach(cp); 2505 g_destroy_consumer(cp); 2506 g_destroy_geom(gp); 2507 if (error != 0) 2508 return (NULL); 2509 gp = NULL; 2510 2511 if (md.md_provider[0] != '\0' && strcmp(md.md_provider, pp->name) != 0) 2512 return (NULL); 2513 if (md.md_provsize != 0 && md.md_provsize != pp->mediasize) 2514 return (NULL); 2515 if (g_journal_debug >= 2) 2516 journal_metadata_dump(&md); 2517 2518 gp = g_journal_create(mp, pp, &md); 2519 return (gp); 2520} 2521 2522static struct g_journal_softc * 2523g_journal_find_device(struct g_class *mp, const char *name) 2524{ 2525 struct g_journal_softc *sc; 2526 struct g_geom *gp; 2527 struct g_provider *pp; 2528 2529 if (strncmp(name, "/dev/", 5) == 0) 2530 name += 5; 2531 LIST_FOREACH(gp, &mp->geom, geom) { 2532 sc = gp->softc; 2533 if (sc == NULL) 2534 continue; 2535 if (sc->sc_flags & GJF_DEVICE_DESTROY) 2536 continue; 2537 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) 2538 continue; 2539 pp = LIST_FIRST(&gp->provider); 2540 if (strcmp(sc->sc_name, name) == 0) 2541 return (sc); 2542 if (pp != NULL && strcmp(pp->name, name) == 0) 2543 return (sc); 2544 } 2545 return (NULL); 2546} 2547 2548static void 2549g_journal_ctl_destroy(struct gctl_req *req, struct g_class *mp) 2550{ 2551 struct g_journal_softc *sc; 2552 const char *name; 2553 char param[16]; 2554 int *nargs; 2555 int error, i; 2556 2557 g_topology_assert(); 2558 2559 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 2560 if (nargs == NULL) { 2561 gctl_error(req, "No '%s' argument.", "nargs"); 2562 return; 2563 } 2564 if (*nargs <= 0) { 2565 gctl_error(req, "Missing device(s)."); 2566 return; 2567 } 2568 2569 for (i = 0; i < *nargs; i++) { 2570 snprintf(param, sizeof(param), "arg%d", i); 2571 name = gctl_get_asciiparam(req, param); 2572 if (name == NULL) { 2573 gctl_error(req, "No 'arg%d' argument.", i); 2574 return; 2575 } 2576 sc = g_journal_find_device(mp, name); 2577 if (sc == NULL) { 2578 gctl_error(req, "No such device: %s.", name); 2579 return; 2580 } 2581 error = g_journal_destroy(sc); 2582 if (error != 0) { 2583 gctl_error(req, "Cannot destroy device %s (error=%d).", 2584 LIST_FIRST(&sc->sc_geom->provider)->name, error); 2585 return; 2586 } 2587 } 2588} 2589 2590static void 2591g_journal_ctl_sync(struct gctl_req *req __unused, struct g_class *mp __unused) 2592{ 2593 2594 g_topology_assert(); 2595 g_topology_unlock(); 2596 g_journal_sync_requested++; 2597 wakeup(&g_journal_switcher_state); 2598 while (g_journal_sync_requested > 0) 2599 tsleep(&g_journal_sync_requested, PRIBIO, "j:sreq", hz / 2); 2600 g_topology_lock(); 2601} 2602 2603static void 2604g_journal_config(struct gctl_req *req, struct g_class *mp, const char *verb) 2605{ 2606 uint32_t *version; 2607 2608 g_topology_assert(); 2609 2610 version = gctl_get_paraml(req, "version", sizeof(*version)); 2611 if (version == NULL) { 2612 gctl_error(req, "No '%s' argument.", "version"); 2613 return; 2614 } 2615 if (*version != G_JOURNAL_VERSION) { 2616 gctl_error(req, "Userland and kernel parts are out of sync."); 2617 return; 2618 } 2619 2620 if (strcmp(verb, "destroy") == 0 || strcmp(verb, "stop") == 0) { 2621 g_journal_ctl_destroy(req, mp); 2622 return; 2623 } else if (strcmp(verb, "sync") == 0) { 2624 g_journal_ctl_sync(req, mp); 2625 return; 2626 } 2627 2628 gctl_error(req, "Unknown verb."); 2629} 2630 2631static void 2632g_journal_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, 2633 struct g_consumer *cp, struct g_provider *pp) 2634{ 2635 struct g_journal_softc *sc; 2636 2637 g_topology_assert(); 2638 2639 sc = gp->softc; 2640 if (sc == NULL) 2641 return; 2642 if (pp != NULL) { 2643 /* Nothing here. */ 2644 } else if (cp != NULL) { 2645 int first = 1; 2646 2647 sbuf_printf(sb, "%s<Role>", indent); 2648 if (cp == sc->sc_dconsumer) { 2649 sbuf_printf(sb, "Data"); 2650 first = 0; 2651 } 2652 if (cp == sc->sc_jconsumer) { 2653 if (!first) 2654 sbuf_printf(sb, ","); 2655 sbuf_printf(sb, "Journal"); 2656 } 2657 sbuf_printf(sb, "</Role>\n"); 2658 if (cp == sc->sc_jconsumer) { 2659 sbuf_printf(sb, "<Jstart>%jd</Jstart>", 2660 (intmax_t)sc->sc_jstart); 2661 sbuf_printf(sb, "<Jend>%jd</Jend>", 2662 (intmax_t)sc->sc_jend); 2663 } 2664 } else { 2665 sbuf_printf(sb, "%s<ID>%u</ID>\n", indent, (u_int)sc->sc_id); 2666 } 2667} 2668 2669static eventhandler_tag g_journal_event_shutdown = NULL; 2670static eventhandler_tag g_journal_event_lowmem = NULL; 2671 2672static void 2673g_journal_shutdown(void *arg, int howto __unused) 2674{ 2675 struct g_class *mp; 2676 struct g_geom *gp, *gp2; 2677 2678 if (panicstr != NULL) 2679 return; 2680 mp = arg; 2681 DROP_GIANT(); 2682 g_topology_lock(); 2683 LIST_FOREACH_SAFE(gp, &mp->geom, geom, gp2) { 2684 if (gp->softc == NULL) 2685 continue; 2686 GJ_DEBUG(0, "Shutting down geom %s.", gp->name); 2687 g_journal_destroy(gp->softc); 2688 } 2689 g_topology_unlock(); 2690 PICKUP_GIANT(); 2691} 2692 2693/* 2694 * Free cached requests from inactive queue in case of low memory. 2695 * We free GJ_FREE_AT_ONCE elements at once. 2696 */ 2697#define GJ_FREE_AT_ONCE 4 2698static void 2699g_journal_lowmem(void *arg, int howto __unused) 2700{ 2701 struct g_journal_softc *sc; 2702 struct g_class *mp; 2703 struct g_geom *gp; 2704 struct bio *bp; 2705 u_int nfree = GJ_FREE_AT_ONCE; 2706 2707 g_journal_stats_low_mem++; 2708 mp = arg; 2709 DROP_GIANT(); 2710 g_topology_lock(); 2711 LIST_FOREACH(gp, &mp->geom, geom) { 2712 sc = gp->softc; 2713 if (sc == NULL || (sc->sc_flags & GJF_DEVICE_DESTROY)) 2714 continue; 2715 mtx_lock(&sc->sc_mtx); 2716 for (bp = sc->sc_inactive.jj_queue; nfree > 0 && bp != NULL; 2717 nfree--, bp = bp->bio_next) { 2718 /* 2719 * This is safe to free the bio_data, because: 2720 * 1. If bio_data is NULL it will be read from the 2721 * inactive journal. 2722 * 2. If bp is sent down, it is first removed from the 2723 * inactive queue, so it's impossible to free the 2724 * data from under in-flight bio. 2725 * On the other hand, freeing elements from the active 2726 * queue, is not safe. 2727 */ 2728 if (bp->bio_data != NULL) { 2729 GJ_DEBUG(2, "Freeing data from %s.", 2730 sc->sc_name); 2731 gj_free(bp->bio_data, bp->bio_length); 2732 bp->bio_data = NULL; 2733 } 2734 } 2735 mtx_unlock(&sc->sc_mtx); 2736 if (nfree == 0) 2737 break; 2738 } 2739 g_topology_unlock(); 2740 PICKUP_GIANT(); 2741} 2742 2743static void g_journal_switcher(void *arg); 2744 2745static void 2746g_journal_init(struct g_class *mp) 2747{ 2748 int error; 2749 2750 /* Pick a conservative value if provided value sucks. */ 2751 if (g_journal_cache_divisor <= 0 || 2752 (vm_kmem_size / g_journal_cache_divisor == 0)) { 2753 g_journal_cache_divisor = 5; 2754 } 2755 if (g_journal_cache_limit > 0) { 2756 g_journal_cache_limit = vm_kmem_size / g_journal_cache_divisor; 2757 g_journal_cache_low = 2758 (g_journal_cache_limit / 100) * g_journal_cache_switch; 2759 } 2760 g_journal_event_shutdown = EVENTHANDLER_REGISTER(shutdown_post_sync, 2761 g_journal_shutdown, mp, EVENTHANDLER_PRI_FIRST); 2762 if (g_journal_event_shutdown == NULL) 2763 GJ_DEBUG(0, "Warning! Cannot register shutdown event."); 2764 g_journal_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, 2765 g_journal_lowmem, mp, EVENTHANDLER_PRI_FIRST); 2766 if (g_journal_event_lowmem == NULL) 2767 GJ_DEBUG(0, "Warning! Cannot register lowmem event."); 2768 error = kthread_create(g_journal_switcher, mp, NULL, 0, 0, 2769 "g_journal switcher"); 2770 KASSERT(error == 0, ("Cannot create switcher thread.")); 2771} 2772 2773static void 2774g_journal_fini(struct g_class *mp) 2775{ 2776 2777 if (g_journal_event_shutdown != NULL) { 2778 EVENTHANDLER_DEREGISTER(shutdown_post_sync, 2779 g_journal_event_shutdown); 2780 } 2781 if (g_journal_event_lowmem != NULL) 2782 EVENTHANDLER_DEREGISTER(vm_lowmem, g_journal_event_lowmem); 2783 g_journal_switcher_state = GJ_SWITCHER_DIE; 2784 wakeup(&g_journal_switcher_state); 2785 while (g_journal_switcher_state != GJ_SWITCHER_DIED) 2786 tsleep(&g_journal_switcher_state, PRIBIO, "jfini:wait", hz / 5); 2787 GJ_DEBUG(1, "Switcher died."); 2788} 2789 2790DECLARE_GEOM_CLASS(g_journal_class, g_journal); 2791 2792static const struct g_journal_desc * 2793g_journal_find_desc(const char *fstype) 2794{ 2795 const struct g_journal_desc *desc; 2796 int i; 2797 2798 for (desc = g_journal_filesystems[i = 0]; desc != NULL; 2799 desc = g_journal_filesystems[++i]) { 2800 if (strcmp(desc->jd_fstype, fstype) == 0) 2801 break; 2802 } 2803 return (desc); 2804} 2805 2806static void 2807g_journal_switch_wait(struct g_journal_softc *sc) 2808{ 2809 struct bintime bt; 2810 2811 mtx_assert(&sc->sc_mtx, MA_OWNED); 2812 if (g_journal_debug >= 2) { 2813 if (sc->sc_flush_in_progress > 0) { 2814 GJ_DEBUG(2, "%d requests flushing.", 2815 sc->sc_flush_in_progress); 2816 } 2817 if (sc->sc_copy_in_progress > 0) { 2818 GJ_DEBUG(2, "%d requests copying.", 2819 sc->sc_copy_in_progress); 2820 } 2821 if (sc->sc_flush_count > 0) { 2822 GJ_DEBUG(2, "%d requests to flush.", 2823 sc->sc_flush_count); 2824 } 2825 if (sc->sc_delayed_count > 0) { 2826 GJ_DEBUG(2, "%d requests delayed.", 2827 sc->sc_delayed_count); 2828 } 2829 } 2830 g_journal_stats_switches++; 2831 if (sc->sc_copy_in_progress > 0) 2832 g_journal_stats_wait_for_copy++; 2833 GJ_TIMER_START(1, &bt); 2834 sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH; 2835 sc->sc_flags |= GJF_DEVICE_SWITCH; 2836 wakeup(sc); 2837 while (sc->sc_flags & GJF_DEVICE_SWITCH) { 2838 msleep(&sc->sc_journal_copying, &sc->sc_mtx, PRIBIO, 2839 "gj:switch", 0); 2840 } 2841 GJ_TIMER_STOP(1, &bt, "Switch time of %s", sc->sc_name); 2842} 2843 2844static void 2845g_journal_do_switch(struct g_class *classp, struct thread *td) 2846{ 2847 struct g_journal_softc *sc; 2848 const struct g_journal_desc *desc; 2849 struct g_geom *gp; 2850 struct mount *mp; 2851 struct bintime bt; 2852 char *mountpoint; 2853 int asyncflag, error, vfslocked; 2854 2855 DROP_GIANT(); 2856 g_topology_lock(); 2857 LIST_FOREACH(gp, &classp->geom, geom) { 2858 sc = gp->softc; 2859 if (sc == NULL) 2860 continue; 2861 if (sc->sc_flags & GJF_DEVICE_DESTROY) 2862 continue; 2863 if ((sc->sc_type & GJ_TYPE_COMPLETE) != GJ_TYPE_COMPLETE) 2864 continue; 2865 mtx_lock(&sc->sc_mtx); 2866 sc->sc_flags |= GJF_DEVICE_BEFORE_SWITCH; 2867 mtx_unlock(&sc->sc_mtx); 2868 } 2869 g_topology_unlock(); 2870 PICKUP_GIANT(); 2871 2872 mtx_lock(&mountlist_mtx); 2873 TAILQ_FOREACH(mp, &mountlist, mnt_list) { 2874 if (mp->mnt_gjprovider == NULL) 2875 continue; 2876 if (mp->mnt_flag & MNT_RDONLY) 2877 continue; 2878 desc = g_journal_find_desc(mp->mnt_stat.f_fstypename); 2879 if (desc == NULL) 2880 continue; 2881 if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) 2882 continue; 2883 /* mtx_unlock(&mountlist_mtx) was done inside vfs_busy() */ 2884 2885 DROP_GIANT(); 2886 g_topology_lock(); 2887 sc = g_journal_find_device(classp, mp->mnt_gjprovider); 2888 g_topology_unlock(); 2889 PICKUP_GIANT(); 2890 2891 if (sc == NULL) { 2892 GJ_DEBUG(0, "Cannot find journal geom for %s.", 2893 mp->mnt_gjprovider); 2894 goto next; 2895 } else if (JEMPTY(sc)) { 2896 mtx_lock(&sc->sc_mtx); 2897 sc->sc_flags &= ~GJF_DEVICE_BEFORE_SWITCH; 2898 mtx_unlock(&sc->sc_mtx); 2899 GJ_DEBUG(3, "No need for %s switch.", sc->sc_name); 2900 goto next; 2901 } 2902 2903 mountpoint = mp->mnt_stat.f_mntonname; 2904 2905 vfslocked = VFS_LOCK_GIANT(mp); 2906 2907 error = vn_start_write(NULL, &mp, V_WAIT); 2908 if (error != 0) { 2909 VFS_UNLOCK_GIANT(vfslocked); 2910 GJ_DEBUG(0, "vn_start_write(%s) failed (error=%d).", 2911 mountpoint, error); 2912 goto next; 2913 } 2914 asyncflag = mp->mnt_flag & MNT_ASYNC; 2915 mp->mnt_flag &= ~MNT_ASYNC; 2916 2917 GJ_TIMER_START(1, &bt); 2918 vfs_msync(mp, MNT_NOWAIT); 2919 GJ_TIMER_STOP(1, &bt, "Msync time of %s", mountpoint); 2920 2921 GJ_TIMER_START(1, &bt); 2922 error = VFS_SYNC(mp, MNT_NOWAIT, curthread); 2923 if (error == 0) 2924 GJ_TIMER_STOP(1, &bt, "Sync time of %s", mountpoint); 2925 else { 2926 GJ_DEBUG(0, "Cannot sync file system %s (error=%d).", 2927 mountpoint, error); 2928 } 2929 mp->mnt_flag |= asyncflag; 2930 2931 vn_finished_write(mp); 2932 2933 if (error != 0) { 2934 VFS_UNLOCK_GIANT(vfslocked); 2935 goto next; 2936 } 2937 2938 /* 2939 * Send BIO_FLUSH before freezing the file system, so it can be 2940 * faster after the freeze. 2941 */ 2942 GJ_TIMER_START(1, &bt); 2943 g_journal_flush_cache(sc); 2944 GJ_TIMER_STOP(1, &bt, "BIO_FLUSH time of %s", sc->sc_name); 2945 2946 GJ_TIMER_START(1, &bt); 2947 error = vfs_write_suspend(mp); 2948 VFS_UNLOCK_GIANT(vfslocked); 2949 GJ_TIMER_STOP(1, &bt, "Suspend time of %s", mountpoint); 2950 if (error != 0) { 2951 GJ_DEBUG(0, "Cannot suspend file system %s (error=%d).", 2952 mountpoint, error); 2953 goto next; 2954 } 2955 2956 error = desc->jd_clean(mp); 2957 if (error != 0) 2958 goto next; 2959 2960 mtx_lock(&sc->sc_mtx); 2961 g_journal_switch_wait(sc); 2962 mtx_unlock(&sc->sc_mtx); 2963 2964 vfs_write_resume(mp); 2965next: 2966 mtx_lock(&mountlist_mtx); 2967 vfs_unbusy(mp, td); 2968 } 2969 mtx_unlock(&mountlist_mtx); 2970 2971 sc = NULL; 2972 for (;;) { 2973 DROP_GIANT(); 2974 g_topology_lock(); 2975 LIST_FOREACH(gp, &g_journal_class.geom, geom) { 2976 sc = gp->softc; 2977 if (sc == NULL) 2978 continue; 2979 mtx_lock(&sc->sc_mtx); 2980 if ((sc->sc_type & GJ_TYPE_COMPLETE) == GJ_TYPE_COMPLETE && 2981 !(sc->sc_flags & GJF_DEVICE_DESTROY) && 2982 (sc->sc_flags & GJF_DEVICE_BEFORE_SWITCH)) { 2983 break; 2984 } 2985 mtx_unlock(&sc->sc_mtx); 2986 sc = NULL; 2987 } 2988 g_topology_unlock(); 2989 PICKUP_GIANT(); 2990 if (sc == NULL) 2991 break; 2992 mtx_assert(&sc->sc_mtx, MA_OWNED); 2993 g_journal_switch_wait(sc); 2994 mtx_unlock(&sc->sc_mtx); 2995 } 2996} 2997 2998/* 2999 * TODO: Switcher thread should be started on first geom creation and killed on 3000 * last geom destruction. 3001 */ 3002static void 3003g_journal_switcher(void *arg) 3004{ 3005 struct thread *td = curthread; 3006 struct g_class *mp; 3007 struct bintime bt; 3008 int error; 3009 3010 mp = arg; 3011 for (;;) { 3012 g_journal_switcher_wokenup = 0; 3013 error = tsleep(&g_journal_switcher_state, PRIBIO, "jsw:wait", 3014 g_journal_switch_time * hz); 3015 if (g_journal_switcher_state == GJ_SWITCHER_DIE) { 3016 g_journal_switcher_state = GJ_SWITCHER_DIED; 3017 GJ_DEBUG(1, "Switcher exiting."); 3018 wakeup(&g_journal_switcher_state); 3019 kthread_exit(0); 3020 } 3021 if (error == 0 && g_journal_sync_requested == 0) { 3022 GJ_DEBUG(1, "Out of cache, force switch (used=%u " 3023 "limit=%u).", g_journal_cache_used, 3024 g_journal_cache_limit); 3025 } 3026 GJ_TIMER_START(1, &bt); 3027 g_journal_do_switch(mp, td); 3028 GJ_TIMER_STOP(1, &bt, "Entire switch time"); 3029 if (g_journal_sync_requested > 0) { 3030 g_journal_sync_requested = 0; 3031 wakeup(&g_journal_sync_requested); 3032 } 3033 } 3034} 3035